[V1][Spec Decode] Update N-gram Proposer Interface (#15750)

Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>

.buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh

@@ -10,15 +10,24 @@ set -x
 set -o pipefail
 
 check_gpus() {
-  # check the number of GPUs and GPU type.
-  declare -g gpu_count=$(nvidia-smi --list-gpus | wc -l)
+  if command -v nvidia-smi; then
+    # check the number of GPUs and GPU type.
+    declare -g gpu_count=$(nvidia-smi --list-gpus | wc -l)
+  elif command -v amd-smi; then
+    declare -g gpu_count=$(amd-smi list | grep 'GPU' | wc -l)
+  fi
+
   if [[ $gpu_count -gt 0 ]]; then
     echo "GPU found."
   else
     echo "Need at least 1 GPU to run benchmarking."
     exit 1
   fi
-  declare -g gpu_type=$(nvidia-smi --query-gpu=name --format=csv,noheader | awk '{print $2}')
+  if command -v nvidia-smi; then
+    declare -g gpu_type=$(nvidia-smi --query-gpu=name --format=csv,noheader | awk '{print $2}')
+  elif command -v amd-smi; then
+    declare -g gpu_type=$(amd-smi static -g 0 -a | grep 'MARKET_NAME' | awk '{print $2}')
+  fi
   echo "GPU type is $gpu_type"
 }
 
@@ -90,9 +99,15 @@ kill_gpu_processes() {
 
 
   # wait until GPU memory usage smaller than 1GB
-  while [ "$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits | head -n 1)" -ge 1000 ]; do
-    sleep 1
-  done
+  if command -v nvidia-smi; then
+    while [ "$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits | head -n 1)" -ge 1000 ]; do
+      sleep 1
+    done
+  elif command -v amd-smi; then
+    while [ "$(amd-smi metric -g 0 | grep 'USED_VRAM' | awk '{print $2}')" -ge 1000 ]; do
+      sleep 1
+    done
+  fi
 
   # remove vllm config file
   rm -rf ~/.config/vllm
@@ -361,7 +376,7 @@ main() {
   # get the current IP address, required by benchmark_serving.py
   export VLLM_HOST_IP=$(hostname -I | awk '{print $1}')
   # turn of the reporting of the status of each request, to clean up the terminal output
-  export VLLM_LOG_LEVEL="WARNING"
+  export VLLM_LOGGING_LEVEL="WARNING"
 
   # prepare for benchmarking
   cd benchmarks || exit 1

.buildkite/nightly-benchmarks/tests/serving-tests.json

@@ -63,10 +63,12 @@
             "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",
             "disable_log_requests": "", 
             "tensor_parallel_size": 4,
-            "swap_space": 16, 
-            "speculative_model": "turboderp/Qwama-0.5B-Instruct",
-            "num_speculative_tokens": 4,
-            "speculative_draft_tensor_parallel_size": 1
+            "swap_space": 16,
+            "speculative_config": {
+                "model": "turboderp/Qwama-0.5B-Instruct",
+                "num_speculative_tokens": 4,
+                "draft_tensor_parallel_size": 1
+            }
         },
         "client_parameters": {
             "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",

.buildkite/release-pipeline.yaml

@@ -3,10 +3,10 @@ steps:
     agents:
       queue: cpu_queue_postmerge
     commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.4.0 --tag vllm-ci:build-image --target build --progress plain ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.4.0 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
       - "mkdir artifacts"
       - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
-      - "bash .buildkite/upload-wheels.sh"
+      - "bash .buildkite/scripts/upload-wheels.sh"
     env:
       DOCKER_BUILDKIT: "1"
 
@@ -14,10 +14,10 @@ steps:
     agents:
       queue: cpu_queue_postmerge
     commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.1.0 --tag vllm-ci:build-image --target build --progress plain ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.1.0 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
       - "mkdir artifacts"
       - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
-      - "bash .buildkite/upload-wheels.sh"
+      - "bash .buildkite/scripts/upload-wheels.sh"
     env:
       DOCKER_BUILDKIT: "1"
 
@@ -31,10 +31,10 @@ steps:
     agents:
       queue: cpu_queue_postmerge
     commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=11.8.0 --tag vllm-ci:build-image --target build --progress plain ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=11.8.0 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
       - "mkdir artifacts"
       - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
-      - "bash .buildkite/upload-wheels.sh"
+      - "bash .buildkite/scripts/upload-wheels.sh"
     env:
       DOCKER_BUILDKIT: "1"
 
@@ -48,7 +48,7 @@ steps:
       queue: cpu_queue_postmerge
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.4.0 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT --target vllm-openai --progress plain ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.4.0 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT --target vllm-openai --progress plain -f docker/Dockerfile ."
       - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
 
   - label: "Build and publish TPU release image"
@@ -57,7 +57,7 @@ steps:
     agents:
       queue: tpu_queue_postmerge
     commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --tag vllm/vllm-tpu:nightly --tag vllm/vllm-tpu:$BUILDKITE_COMMIT --progress plain -f Dockerfile.tpu ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --tag vllm/vllm-tpu:nightly --tag vllm/vllm-tpu:$BUILDKITE_COMMIT --progress plain -f docker/Dockerfile.tpu ."
       - "docker push vllm/vllm-tpu:nightly"
       - "docker push vllm/vllm-tpu:$BUILDKITE_COMMIT"
     plugins:
@@ -82,7 +82,7 @@ steps:
       queue: cpu_queue_postmerge
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --progress plain -f Dockerfile.cpu ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest --progress plain --target vllm-openai -f docker/Dockerfile.cpu ."
       - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version)"
     env:
       DOCKER_BUILDKIT: "1"

.buildkite/run-openvino-test.sh

@@ -1,16 +0,0 @@
-#!/bin/bash
-
-# This script build the OpenVINO docker image and run the offline inference inside the container.
-# It serves a sanity check for compilation and basic model usage.
-set -ex
-
-# Try building the docker image
-docker build -t openvino-test -f Dockerfile.openvino .
-
-# Setup cleanup
-remove_docker_container() { docker rm -f openvino-test || true; }
-trap remove_docker_container EXIT
-remove_docker_container
-
-# Run the image and launch offline inference
-docker run --network host --env VLLM_OPENVINO_KVCACHE_SPACE=1 --name openvino-test openvino-test python3 /workspace/examples/offline_inference/basic/generate.py --model facebook/opt-125m

.buildkite/run-tpu-test.sh

@@ -1,25 +0,0 @@
-#!/bin/bash
-
-set -e
-
-# Build the docker image.
-docker build -f Dockerfile.tpu -t vllm-tpu .
-
-# Set up cleanup.
-remove_docker_container() { docker rm -f tpu-test || true; }
-trap remove_docker_container EXIT
-# Remove the container that might not be cleaned up in the previous run.
-remove_docker_container
-
-# For HF_TOKEN.
-source /etc/environment
-# Run a simple end-to-end example.
-docker run --privileged --net host --shm-size=16G -it \
-    -e "HF_TOKEN=$HF_TOKEN" --name tpu-test \
-    vllm-tpu /bin/bash -c "python3 -m pip install git+https://github.com/thuml/depyf.git \
-    && python3 -m pip install pytest \
-    && python3 -m pip install lm_eval[api]==0.4.4 \
-    && pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py \
-    && python3 /workspace/vllm/tests/tpu/test_compilation.py \
-    && python3 /workspace/vllm/tests/tpu/test_quantization_accuracy.py \
-    && python3 /workspace/vllm/examples/offline_inference/tpu.py"

.buildkite/run-tpu-v1-test.sh

@@ -1,27 +0,0 @@
-#!/bin/bash
-
-set -e
-
-# Build the docker image.
-docker build -f Dockerfile.tpu -t vllm-tpu .
-
-# Set up cleanup.
-remove_docker_container() { docker rm -f tpu-test || true; }
-trap remove_docker_container EXIT
-# Remove the container that might not be cleaned up in the previous run.
-remove_docker_container
-
-# For HF_TOKEN.
-source /etc/environment
-# Run a simple end-to-end example.
-docker run --privileged --net host --shm-size=16G -it \
-    -e "HF_TOKEN=$HF_TOKEN" -e "VLLM_USE_V1=1" --name tpu-test \
-    vllm-tpu /bin/bash -c "python3 -m pip install git+https://github.com/thuml/depyf.git \
-    && python3 -m pip install pytest \
-    && python3 -m pip install lm_eval[api]==0.4.4 \
-    && pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py \
-    && pytest -v -s /workspace/vllm/tests/v1/tpu/test_basic.py \
-    && pytest -v -s /workspace/vllm/tests/entrypoints/llm/test_accuracy.py::test_lm_eval_accuracy_v1_engine \
-    && python3 /workspace/vllm/tests/tpu/test_compilation.py \
-    && python3 /workspace/vllm/tests/tpu/test_quantization_accuracy.py \
-    && python3 /workspace/vllm/examples/offline_inference/tpu.py"

.buildkite/scripts/hardware_ci/run-amd-test.sh

@@ -105,19 +105,33 @@ fi
 if [[ $commands == *" entrypoints/openai "* ]]; then
   commands=${commands//" entrypoints/openai "/" entrypoints/openai \
   --ignore=entrypoints/openai/test_audio.py \
-  --ignore=entrypoints/openai/test_chat.py \
   --ignore=entrypoints/openai/test_shutdown.py \
   --ignore=entrypoints/openai/test_completion.py \
   --ignore=entrypoints/openai/test_sleep.py \
   --ignore=entrypoints/openai/test_models.py \
+  --ignore=entrypoints/openai/test_lora_adapters.py \
+  --ignore=entrypoints/openai/test_return_tokens_as_ids.py \
+  --ignore=entrypoints/openai/test_root_path.py \
+  --ignore=entrypoints/openai/test_tokenization.py \
   --ignore=entrypoints/openai/test_prompt_validation.py "}
 fi
 
 #ignore certain Entrypoints/llm tests
-if [[ $commands == *" && pytest -v -s entrypoints/llm/test_guided_generate.py"* ]]; then
-  commands=${commands//" && pytest -v -s entrypoints/llm/test_guided_generate.py"/" "}
+if [[ $commands == *" entrypoints/llm "* ]]; then
+  commands=${commands//" entrypoints/llm "/" entrypoints/llm \
+  --ignore=entrypoints/llm/test_chat.py \
+  --ignore=entrypoints/llm/test_accuracy.py \
+  --ignore=entrypoints/llm/test_init.py \
+  --ignore=entrypoints/llm/test_generate_multiple_loras.py \
+  --ignore=entrypoints/llm/test_prompt_validation.py "}
 fi
 
+#Obsolete currently
+##ignore certain Entrypoints/llm tests
+#if [[ $commands == *" && pytest -v -s entrypoints/llm/test_guided_generate.py"* ]]; then
+#  commands=${commands//" && pytest -v -s entrypoints/llm/test_guided_generate.py"/" "}
+#fi
+
 # --ignore=entrypoints/openai/test_encoder_decoder.py \
 # --ignore=entrypoints/openai/test_embedding.py \
 # --ignore=entrypoints/openai/test_oot_registration.py
@@ -134,9 +148,10 @@ if [[ $commands == *"--shard-id="* ]]; then
     # assign shard-id for each shard
     commands_gpu=${commands//"--shard-id= "/"--shard-id=${GPU} "}
     echo "Shard ${GPU} commands:$commands_gpu"
+    echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
     docker run \
-        --device /dev/kfd --device /dev/dri \
-        --network host \
+        --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
+        --network=host \
         --shm-size=16gb \
         --rm \
         -e HIP_VISIBLE_DEVICES="${GPU}" \
@@ -163,9 +178,10 @@ if [[ $commands == *"--shard-id="* ]]; then
     fi
   done
 else
+  echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
   docker run \
-          --device /dev/kfd --device /dev/dri \
-          --network host \
+          --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
+          --network=host \
           --shm-size=16gb \
           --rm \
           -e HIP_VISIBLE_DEVICES=0 \

.buildkite/scripts/hardware_ci/run-cpu-test-ppc64le.sh

@@ -10,5 +10,5 @@ trap remove_docker_container EXIT
 remove_docker_container
 
 # Try building the docker image
-docker build -t cpu-test -f Dockerfile.ppc64le .
+docker build -t cpu-test -f docker/Dockerfile.ppc64le .
 

.buildkite/scripts/hardware_ci/run-cpu-test.sh

@@ -8,15 +8,19 @@ set -ex
 CORE_RANGE=${CORE_RANGE:-48-95}
 NUMA_NODE=${NUMA_NODE:-1}
 
-# Try building the docker image
-numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build -t cpu-test-"$BUILDKITE_BUILD_NUMBER" -f Dockerfile.cpu .
-numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" -t cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2 -f Dockerfile.cpu .
-
 # Setup cleanup
-remove_docker_container() { set -e; docker rm -f cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" || true; }
+remove_docker_container() { 
+    set -e; 
+    docker rm -f cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" || true; 
+    docker image rm cpu-test-"$BUILDKITE_BUILD_NUMBER" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2 || true; 
+}
 trap remove_docker_container EXIT
 remove_docker_container
 
+# Try building the docker image
+numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --tag cpu-test-"$BUILDKITE_BUILD_NUMBER" --target vllm-test -f docker/Dockerfile.cpu .
+numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" --tag cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2 --target vllm-test -f docker/Dockerfile.cpu .
+
 # Run the image, setting --shm-size=4g for tensor parallel.
 docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus="$CORE_RANGE"  \
  --cpuset-mems="$NUMA_NODE" --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"
@@ -36,8 +40,8 @@ function cpu_tests() {
   # Run basic model test
   docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
     set -e
-    pip install -r vllm/requirements/test.txt
-    pip install -r vllm/requirements/cpu.txt
+    pytest -v -s tests/kernels/test_cache.py -m cpu_model
+    pytest -v -s tests/kernels/test_mla_decode_cpu.py -m cpu_model
     pytest -v -s tests/models/decoder_only/language -m cpu_model
     pytest -v -s tests/models/embedding/language -m cpu_model
     pytest -v -s tests/models/encoder_decoder/language -m cpu_model

.buildkite/scripts/hardware_ci/run-gh200-test.sh

@@ -9,11 +9,13 @@ python3 use_existing_torch.py
 
 # Try building the docker image
 DOCKER_BUILDKIT=1 docker build . \
+  --file docker/Dockerfile \
   --target vllm-openai \
   --platform "linux/arm64" \
   -t gh200-test \
   --build-arg max_jobs=66 \
   --build-arg nvcc_threads=2 \
+  --build-arg RUN_WHEEL_CHECK=false \
   --build-arg torch_cuda_arch_list="9.0+PTX" \
   --build-arg vllm_fa_cmake_gpu_arches="90-real"
 
@@ -23,6 +25,6 @@ trap remove_docker_container EXIT
 remove_docker_container
 
 # Run the image and test offline inference
-docker run -e HF_TOKEN -v /root/.cache/huggingface:/root/.cache/huggingface --name gh200-test --gpus=all --entrypoint="" gh200-test bash -c '
+docker run -e HF_TOKEN -e VLLM_WORKER_MULTIPROC_METHOD=spawn -v /root/.cache/huggingface:/root/.cache/huggingface --name gh200-test --gpus=all --entrypoint="" gh200-test bash -c '
     python3 examples/offline_inference/basic/generate.py --model meta-llama/Llama-3.2-1B
 '

.buildkite/scripts/hardware_ci/run-hpu-test.sh

@@ -5,7 +5,7 @@
 set -ex
 
 # Try building the docker image
-docker build -t hpu-test-env -f Dockerfile.hpu .
+docker build -t hpu-test-env -f docker/Dockerfile.hpu .
 
 # Setup cleanup
 # certain versions of HPU software stack have a bug that can

.buildkite/scripts/hardware_ci/run-neuron-test.sh

@@ -35,7 +35,7 @@ else
     date "+%s" > /tmp/neuron-docker-build-timestamp
 fi
 
-docker build -t "${image_name}" -f Dockerfile.neuron .
+docker build -t "${image_name}" -f docker/Dockerfile.neuron .
 
 # Setup cleanup
 remove_docker_container() {

.buildkite/scripts/hardware_ci/run-tpu-v1-test.sh

@@ -0,0 +1,47 @@
+#!/bin/bash
+
+set -xue
+
+# Build the docker image.
+docker build -f docker/Dockerfile.tpu -t vllm-tpu .
+
+# Set up cleanup.
+remove_docker_container() { docker rm -f tpu-test || true; }
+trap remove_docker_container EXIT
+# Remove the container that might not be cleaned up in the previous run.
+remove_docker_container
+
+# For HF_TOKEN.
+source /etc/environment
+# Run a simple end-to-end example.
+docker run --privileged --net host --shm-size=16G -it \
+    -e "HF_TOKEN=$HF_TOKEN" --name tpu-test \
+    vllm-tpu /bin/bash -c "python3 -m pip install git+https://github.com/thuml/depyf.git \
+    && python3 -m pip install pytest \
+    && python3 -m pip install lm_eval[api]==0.4.4 \
+    && export VLLM_USE_V1=1 \
+    && export VLLM_XLA_CHECK_RECOMPILATION=1 \
+    && echo TEST_0 \
+    && pytest -v -s /workspace/vllm/tests/v1/tpu/test_perf.py \
+    && echo TEST_1 \
+    && pytest -v -s /workspace/vllm/tests/tpu/test_compilation.py \
+    && echo TEST_2 \
+    && pytest -v -s /workspace/vllm/tests/v1/tpu/test_basic.py \
+    && echo TEST_3 \
+    && pytest -v -s /workspace/vllm/tests/entrypoints/llm/test_accuracy.py::test_lm_eval_accuracy_v1_engine \
+    && echo TEST_4 \
+    && pytest -s -v /workspace/vllm/tests/tpu/test_quantization_accuracy.py \
+    && echo TEST_5 \
+    && python3 /workspace/vllm/examples/offline_inference/tpu.py \
+    && echo TEST_6 \
+    && pytest -s -v /workspace/vllm/tests/v1/tpu/worker/test_tpu_model_runner.py \
+    && echo TEST_7 \
+    && pytest -s -v /workspace/vllm/tests/v1/tpu/test_sampler.py \
+    && echo TEST_8 \
+    && pytest -s -v /workspace/vllm/tests/v1/tpu/test_topk_topp_sampler.py \
+    && echo TEST_9 \
+    && pytest -s -v /workspace/vllm/tests/v1/tpu/test_pallas.py" \
+
+
+# TODO: This test fails because it uses RANDOM_SEED sampling
+# && VLLM_USE_V1=1 pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py \

.buildkite/scripts/hardware_ci/run-xpu-test.sh

@@ -8,14 +8,15 @@ image_name="xpu/vllm-ci:${BUILDKITE_COMMIT}"
 container_name="xpu_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
 
 # Try building the docker image
-docker build -t ${image_name} -f Dockerfile.xpu .
+docker build -t ${image_name} -f docker/Dockerfile.xpu .
 
 # Setup cleanup
 remove_docker_container() { 
-  docker rm -f "${container_name}" || docker image rm -f "${image_name}" || true;
+  docker rm -f "${container_name}" || true; 
+  docker image rm -f "${image_name}" || true;
+  docker system prune -f || true;
 }
 trap remove_docker_container EXIT
-remove_docker_container
 
 # Run the image and test offline inference/tensor parallel
 docker run \
@@ -25,6 +26,6 @@ docker run \
     --name "${container_name}" \
     "${image_name}" \
     sh -c '
-    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m
-    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m -tp 2
+    VLLM_USE_V1=0 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m
+    VLLM_USE_V1=0 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m -tp 2
 '

.buildkite/scripts/run-multi-node-test.sh

@@ -3,7 +3,7 @@
 set -euox pipefail
 
 if [[ $# -lt 4 ]]; then
-    echo "Usage: .buildkite/run-multi-node-test.sh WORKING_DIR NUM_NODES NUM_GPUS DOCKER_IMAGE COMMAND1 COMMAND2 ... COMMANDN"
+    echo "Usage: .buildkite/scripts/run-multi-node-test.sh WORKING_DIR NUM_NODES NUM_GPUS DOCKER_IMAGE COMMAND1 COMMAND2 ... COMMANDN"
     exit 1
 fi
 

.buildkite/test-pipeline.yaml

@@ -104,7 +104,7 @@ steps:
 - label: Entrypoints Test # 40min
   working_dir: "/vllm-workspace/tests"
   fast_check: true
-  mirror_hardwares: [amd]
+  #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/
   - tests/entrypoints/llm
@@ -118,7 +118,7 @@ steps:
   - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
   - pytest -v -s entrypoints/llm/test_generate_multiple_loras.py # it needs a clean process
   - VLLM_USE_V1=0 pytest -v -s entrypoints/llm/test_guided_generate.py # it needs a clean process
-  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/correctness/
+  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_oot_registration.py  --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/correctness/
   - pytest -v -s entrypoints/test_chat_utils.py
   - VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
 
@@ -135,8 +135,14 @@ steps:
   - examples/offline_inference/rlhf.py
   - examples/offline_inference/rlhf_colocate.py
   - tests/examples/offline_inference/data_parallel.py
+  - tests/v1/test_async_llm_dp.py
   commands:
+  # test with tp=2 and external_dp=2
+  - VLLM_USE_V1=0 torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
+  - torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
+  # test with internal dp
   - python3 ../examples/offline_inference/data_parallel.py
+  - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
   - pytest -v -s distributed/test_utils.py
   - pytest -v -s compile/test_basic_correctness.py
   - pytest -v -s distributed/test_pynccl.py
@@ -149,6 +155,7 @@ steps:
   - popd
 
 - label: Metrics, Tracing Test # 10min
+  mirror_hardwares: [amd]
   num_gpus: 2
   source_file_dependencies:
   - vllm/
@@ -167,7 +174,7 @@ steps:
 #####  1 GPU test  #####
 
 - label: Regression Test # 5min
-  mirror_hardwares: [amd]
+  #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/
   - tests/test_regression
@@ -198,7 +205,6 @@ steps:
   commands:
     # split the test to avoid interference
     - pytest -v -s v1/core
-    - pytest -v -s v1/entrypoints
     - pytest -v -s v1/engine
     - pytest -v -s v1/entrypoints
     - pytest -v -s v1/sample
@@ -279,11 +285,11 @@ steps:
     - pytest -v -s spec_decode/e2e/test_eagle_correctness.py
 
 - label: LoRA Test %N # 15min each
-  mirror_hardwares: [amd]
+  #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/lora
   - tests/lora
-  command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_long_context.py --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_minicpmv_tp.py  --ignore=lora/test_transfomers_model.py
+  command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py
   parallelism: 4
 
 - label: PyTorch Fullgraph Smoke Test # 9min
@@ -295,6 +301,7 @@ steps:
   # these tests need to be separated, cannot combine
   - pytest -v -s compile/piecewise/test_simple.py
   - pytest -v -s compile/piecewise/test_toy_llama.py
+  - pytest -v -s compile/test_pass_manager.py
 
 - label: PyTorch Fullgraph Test # 18min
   source_file_dependencies:
@@ -304,7 +311,7 @@ steps:
   - pytest -v -s compile/test_full_graph.py
 
 - label: Kernels Test %N # 1h each
-  mirror_hardwares: [amd]
+  # mirror_hardwares: [amd]
   source_file_dependencies:
   - csrc/
   - vllm/attention
@@ -314,7 +321,7 @@ steps:
   parallelism: 4
 
 - label: Tensorizer Test # 11min
-  mirror_hardwares: [amd]
+  # mirror_hardwares: [amd]
   soft_fail: true
   source_file_dependencies:
   - vllm/model_executor/model_loader
@@ -330,7 +337,7 @@ steps:
   source_file_dependencies:
   - benchmarks/
   commands:
-  - bash run-benchmarks.sh
+  - bash scripts/run-benchmarks.sh
 
 - label: Quantization Test # 33min
   source_file_dependencies:
@@ -365,7 +372,7 @@ steps:
 
 - label: OpenAI-Compatible Tool Use # 20 min
   fast_check: false
-  mirror_hardwares: [ amd ]
+  #mirror_hardwares: [ amd ]
   source_file_dependencies:
     - vllm/
     - tests/tool_use
@@ -424,6 +431,7 @@ steps:
     - pytest -v -s models/encoder_decoder/audio_language -m core_model
     - pytest -v -s models/encoder_decoder/language -m core_model
     - pytest -v -s models/encoder_decoder/vision_language -m core_model
+    - pytest -v -s models/decoder_only/vision_language/test_interleaved.py
 
 - label: Multi-Modal Models Test (Extended) 1 # 48m
   optional: true
@@ -456,6 +464,7 @@ steps:
 
 # This test is used only in PR development phase to test individual models and should never run on main
 - label: Custom Models Test
+  mirror_hardwares: [amd]
   optional: true
   commands:
     - echo 'Testing custom models...'
@@ -467,6 +476,7 @@ steps:
 #####  multi gpus test  #####
 
 - label: Distributed Comm Ops Test # 7min
+  mirror_hardwares: [amd]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
@@ -509,10 +519,11 @@ steps:
   - vllm/worker/worker.py
   - vllm/worker/model_runner.py
   - entrypoints/llm/test_collective_rpc.py
+  - tests/v1/test_async_llm_dp.py
+  - vllm/v1/engine/
   commands:
+  - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
   - pytest -v -s entrypoints/llm/test_collective_rpc.py
-  - VLLM_USE_V1=1 torchrun --nproc-per-node=2 distributed/test_torchrun_example.py
-  - torchrun --nproc-per-node=2 distributed/test_torchrun_example.py
   - pytest -v -s ./compile/test_basic_correctness.py
   - pytest -v -s ./compile/test_wrapper.py
   - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
@@ -589,14 +600,10 @@ steps:
     # FIXIT: find out which code initialize cuda before running the test
     # before the fix, we need to use spawn to test it
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    # This test runs llama 13B, so it is required to run on 4 GPUs.
-    - pytest -v -s -x lora/test_long_context.py
     # There is some Tensor Parallelism related processing logic in LoRA that
     # requires multi-GPU testing for validation.
     - pytest -v -s -x lora/test_chatglm3_tp.py
     - pytest -v -s -x lora/test_llama_tp.py
-    - pytest -v -s -x lora/test_minicpmv_tp.py
-    - pytest -v -s -x lora/test_transfomers_model.py
 
 
 - label: Weight Loading Multiple GPU Test  # 33min

.github/ISSUE_TEMPLATE/800-misc-discussion.yml

@@ -1,28 +0,0 @@
-name: 🎲 Misc/random discussions that do not fit into the above categories.
-description: Submit a discussion as you like. Note that developers are heavily overloaded and we mainly rely on community users to answer these issues.
-title: "[Misc]: "
-labels: ["misc"]
-
-body:
-- type: markdown
-  attributes:
-    value: >
-      #### Before submitting an issue, please make sure the issue hasn't been already addressed by searching through [the existing and past issues](https://github.com/vllm-project/vllm/issues?q=is%3Aissue+sort%3Acreated-desc+).
-- type: textarea
-  attributes:
-    label: Anything you want to discuss about vllm.
-    description: >
-      Anything you want to discuss about vllm.
-  validations:
-    required: true
-- type: markdown
-  attributes:
-    value: >
-      Thanks for contributing 🎉!
-- type: checkboxes
-  id: askllm
-  attributes:
-    label: Before submitting a new issue...
-    options:
-      - label: Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions.
-        required: true

.github/ISSUE_TEMPLATE/config.yml

@@ -1 +1,5 @@
 blank_issues_enabled: false
+contact_links:
+  - name: Questions
+    url: https://discuss.vllm.ai
+    about: Ask questions and discuss with other vLLM community members

.github/mergify.yml

@@ -19,7 +19,7 @@ pull_request_rules:
       - files~=\.buildkite/
       - files~=^cmake/
       - files=CMakeLists.txt
-      - files~=^Dockerfile
+      - files~=^docker/Dockerfile
       - files~=^requirements.*\.txt
       - files=setup.py
   actions:
@@ -88,6 +88,36 @@ pull_request_rules:
       add:
         - v1
 
+- name: label-tpu
+  description: Automatically apply tpu label
+  # Keep this list in sync with `label-tpu-remove` conditions
+  conditions:
+    - or:
+      - files~=tpu.py
+      - files~=_tpu
+      - files~=tpu_
+      - files~=/tpu/
+      - files~=pallas
+  actions:
+    label:
+      add:
+        - tpu
+
+- name: label-tpu-remove
+  description: Automatically remove tpu label
+  # Keep this list in sync with `label-tpu` conditions
+  conditions:
+    - and:
+      - -files~=tpu.py
+      - -files~=_tpu
+      - -files~=tpu_
+      - -files~=/tpu/
+      - -files~=pallas
+  actions:
+    label:
+      remove:
+        - tpu
+
 - name: ping author on conflicts and add 'needs-rebase' label
   conditions:
       - conflict

.github/workflows/lint-and-deploy.yaml

@@ -50,7 +50,7 @@ jobs:
         uses: helm/kind-action@a1b0e391336a6ee6713a0583f8c6240d70863de3 # v1.12.0
 
       - name: Build the Docker image vllm cpu
-        run: docker buildx build -f Dockerfile.cpu -t vllm-cpu-env .
+        run: docker buildx build -f docker/Dockerfile.cpu -t vllm-cpu-env .
 
       - name: Configuration of docker images, network and namespace for the kind cluster
         run: |

.gitignore

@@ -2,7 +2,8 @@
 /vllm/_version.py
 
 # vllm-flash-attn built from source
-vllm/vllm_flash_attn/
+vllm/vllm_flash_attn/*
+!vllm/vllm_flash_attn/fa_utils.py
 
 # Byte-compiled / optimized / DLL files
 __pycache__/

.pre-commit-config.yaml

@@ -1,3 +1,6 @@
+default_install_hook_types:
+  - pre-commit
+  - commit-msg
 default_stages:
   - pre-commit # Run locally
   - manual # Run in CI

CMakeLists.txt

@@ -34,7 +34,7 @@ set(PYTHON_SUPPORTED_VERSIONS "3.9" "3.10" "3.11" "3.12")
 set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0")
 
 # Supported AMD GPU architectures.
-set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101")
+set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201")
 
 #
 # Supported/expected torch versions for CUDA/ROCm.
@@ -44,7 +44,7 @@ set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101")
 #
 # Note: the CUDA torch version is derived from pyproject.toml and various
 # requirements.txt files and should be kept consistent.  The ROCm torch
-# versions are derived from Dockerfile.rocm
+# versions are derived from docker/Dockerfile.rocm
 #
 set(TORCH_SUPPORTED_VERSION_CUDA "2.6.0")
 set(TORCH_SUPPORTED_VERSION_ROCM "2.6.0")
@@ -234,6 +234,7 @@ set(VLLM_EXT_SRC
   "csrc/activation_kernels.cu"
   "csrc/layernorm_kernels.cu"
   "csrc/layernorm_quant_kernels.cu"
+  "csrc/cuda_view.cu"
   "csrc/quantization/gptq/q_gemm.cu"
   "csrc/quantization/compressed_tensors/int8_quant_kernels.cu"
   "csrc/quantization/fp8/common.cu"
@@ -241,6 +242,7 @@ set(VLLM_EXT_SRC
   "csrc/quantization/gguf/gguf_kernel.cu"
   "csrc/cuda_utils_kernels.cu"
   "csrc/prepare_inputs/advance_step.cu"
+  "csrc/custom_all_reduce.cu"
   "csrc/torch_bindings.cpp")
 
 if(VLLM_GPU_LANG STREQUAL "CUDA")
@@ -282,7 +284,6 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     "csrc/mamba/causal_conv1d/causal_conv1d.cu"
     "csrc/quantization/aqlm/gemm_kernels.cu"
     "csrc/quantization/awq/gemm_kernels.cu"
-    "csrc/custom_all_reduce.cu"
     "csrc/permute_cols.cu"
     "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu"
     "csrc/quantization/fp4/nvfp4_quant_entry.cu"
@@ -461,6 +462,33 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     set(FP4_ARCHS)
   endif()
 
+  #
+  # CUTLASS MoE kernels
+
+  # The MoE kernel cutlass_moe_mm requires CUDA 12.3 or later (and only works
+  # on Hopper). get_cutlass_moe_mm_data should only be compiled if it's possible
+  # to compile MoE kernels that use its output.
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu"
+             "csrc/quantization/cutlass_w8a8/moe/moe_data.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM90=1")
+    message(STATUS "Building grouped_mm_c3x for archs: ${SCALED_MM_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
+      message(STATUS "Not building grouped_mm_c3x kernels as CUDA Compiler version is "
+                     "not >= 12.3, we recommend upgrading to CUDA 12.3 or later "
+                     "if you intend on running FP8 quantized MoE models on Hopper.")
+    else()
+      message(STATUS "Not building grouped_mm_c3x as no compatible archs found "
+                     "in CUDA target architectures")
+    endif()
+  endif()
+
   #
   # Machete kernels
 

Dockerfile.cpu

@@ -1,69 +0,0 @@
-# This vLLM Dockerfile is used to construct image that can build and run vLLM on x86 CPU platform.
-
-FROM ubuntu:22.04 AS cpu-test-1
-
-ENV CCACHE_DIR=/root/.cache/ccache
-
-ENV CMAKE_CXX_COMPILER_LAUNCHER=ccache
-
-RUN --mount=type=cache,target=/var/cache/apt \
-    apt-get update -y \
-    && apt-get install -y curl ccache git wget vim numactl gcc-12 g++-12 python3 python3-pip libtcmalloc-minimal4 libnuma-dev \
-    && apt-get install -y ffmpeg libsm6 libxext6 libgl1 \
-    && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 10 --slave /usr/bin/g++ g++ /usr/bin/g++-12
-
-# https://intel.github.io/intel-extension-for-pytorch/cpu/latest/tutorials/performance_tuning/tuning_guide.html
-# intel-openmp provides additional performance improvement vs. openmp
-# tcmalloc provides better memory allocation efficiency, e.g, holding memory in caches to speed up access of commonly-used objects.
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install intel-openmp==2025.0.1
-
-ENV LD_PRELOAD="/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:/usr/local/lib/libiomp5.so"
-
-RUN echo 'ulimit -c 0' >> ~/.bashrc
-
-RUN pip install intel_extension_for_pytorch==2.6.0
-
-WORKDIR /workspace
-
-ARG PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu"
-ENV PIP_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
-RUN --mount=type=cache,target=/root/.cache/pip \
-    --mount=type=bind,src=requirements/build.txt,target=requirements/build.txt \
-    pip install --upgrade pip && \
-    pip install -r requirements/build.txt
-
-FROM cpu-test-1 AS build
-
-WORKDIR /workspace/vllm
-
-RUN --mount=type=cache,target=/root/.cache/pip \
-    --mount=type=bind,src=requirements/common.txt,target=requirements/common.txt \
-    --mount=type=bind,src=requirements/cpu.txt,target=requirements/cpu.txt \
-    pip install -v -r requirements/cpu.txt
-
-COPY . .
-ARG GIT_REPO_CHECK=0
-RUN --mount=type=bind,source=.git,target=.git \
-    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh ; fi
-
-# Support for building with non-AVX512 vLLM: docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" ...
-ARG VLLM_CPU_DISABLE_AVX512
-ENV VLLM_CPU_DISABLE_AVX512=${VLLM_CPU_DISABLE_AVX512}
-
-RUN --mount=type=cache,target=/root/.cache/pip \
-    --mount=type=cache,target=/root/.cache/ccache \
-    --mount=type=bind,source=.git,target=.git \
-    VLLM_TARGET_DEVICE=cpu python3 setup.py bdist_wheel && \
-    pip install dist/*.whl && \
-    rm -rf dist
-
-WORKDIR /workspace/
-
-RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
-
-# install development dependencies (for testing)
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install -e tests/vllm_test_utils
-
-ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]

Dockerfile.openvino

@@ -1,29 +0,0 @@
-# The vLLM Dockerfile is used to construct vLLM image that can be directly used
-# to run the OpenAI compatible server.
-
-FROM ubuntu:22.04 AS dev
-
-RUN apt-get update -y && \
-    apt-get install -y \
-        git python3-pip \
-        ffmpeg libsm6 libxext6 libgl1
-WORKDIR /workspace
-
-COPY . .
-ARG GIT_REPO_CHECK=0
-RUN --mount=type=bind,source=.git,target=.git \
-    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh ; fi
-
-RUN python3 -m pip install -U pip
-# install build requirements
-RUN PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu" python3 -m pip install -r /workspace/requirements/build.txt
-# build vLLM with OpenVINO backend
-RUN PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu" VLLM_TARGET_DEVICE="openvino" python3 -m pip install /workspace
-
-COPY examples/ /workspace/examples
-COPY benchmarks/ /workspace/benchmarks
-
-# install development dependencies (for testing)
-RUN python3 -m pip install -e tests/vllm_test_utils
-
-CMD ["/bin/bash"]

Dockerfile.ppc64le

@@ -1,37 +0,0 @@
-FROM mambaorg/micromamba
-ARG MAMBA_DOCKERFILE_ACTIVATE=1
-USER root
-
-ENV PATH="/usr/local/cargo/bin:$PATH:/opt/conda/bin/"
-
-RUN apt-get update -y && apt-get install -y git wget kmod curl vim libnuma-dev libsndfile-dev libprotobuf-dev build-essential ffmpeg libsm6 libxext6 libgl1 libssl-dev 
-
-# Some packages in requirements/cpu are installed here
-# IBM provides optimized packages for ppc64le processors in the open-ce project for mamba
-# Currently these may not be available for venv or pip directly
-RUN micromamba install -y -n base -c https://ftp.osuosl.org/pub/open-ce/1.11.0-p10/ -c defaults python=3.10 rust && micromamba clean --all --yes
-
-COPY ./ /workspace/vllm
-
-WORKDIR /workspace/vllm
-ARG GIT_REPO_CHECK=0
-RUN --mount=type=bind,source=.git,target=.git \
-    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh; fi
-
-RUN --mount=type=cache,target=/root/.cache/pip  \
-    RUSTFLAGS='-L /opt/conda/lib' pip install -v --prefer-binary --extra-index-url https://repo.fury.io/mgiessing \
-        'cmake>=3.26' ninja packaging 'setuptools-scm>=8' wheel jinja2 \
-        -r requirements/cpu.txt \
-        xformers uvloop==0.20.0
-
-RUN --mount=type=bind,source=.git,target=.git \
-    VLLM_TARGET_DEVICE=cpu python3 setup.py install
-
-# install development dependencies (for testing)
-RUN python3 -m pip install -e tests/vllm_test_utils
-
-WORKDIR /workspace/
-
-RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
-
-ENTRYPOINT ["/opt/conda/bin/python3", "-m", "vllm.entrypoints.openai.api_server"]

README.md

@@ -10,17 +10,27 @@ Easy, fast, and cheap LLM serving for everyone
 </h3>
 
 <p align="center">
-| <a href="https://docs.vllm.ai"><b>Documentation</b></a> | <a href="https://vllm.ai"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://x.com/vllm_project"><b>Twitter/X</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
+| <a href="https://docs.vllm.ai"><b>Documentation</b></a> | <a href="https://vllm.ai"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://x.com/vllm_project"><b>Twitter/X</b></a> | <a href="https://discuss.vllm.ai"><b>User Forum</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
 </p>
 
-*Latest News* 🔥
+---
 
-- [2025/03] We hosted [the first vLLM China Meetup](https://mp.weixin.qq.com/s/n77GibL2corAtQHtVEAzfg)! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1REHvfQMKGnvz6p3Fd23HhSO4c8j5WPGZV0bKYLwnHyQ/edit#slide=id.g33fb1ff286e_0_29).
+[2025/04] We're hosting our first-ever *vLLM Asia Developer Day* in Singapore on *April 3rd*! This is a full-day event (9 AM - 9 PM SGT) in partnership with SGInnovate, AMD, and Embedded LLM. Meet the vLLM team and learn about LLM inference for RL, MI300X, and more! [Register Now](https://www.sginnovate.com/event/limited-availability-morning-evening-slots-remaining-inaugural-vllm-asia-developer-day)
+
+---
+
+*Latest News* 🔥
+- [2025/03] We hosted [vLLM x Ollama Inference Night](https://lu.ma/vllm-ollama)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/16T2PDD1YwRnZ4Tu8Q5r6n53c5Lr5c73UV9Vd2_eBo4U/edit?usp=sharing).
+- [2025/03] We hosted [the first vLLM China Meetup](https://mp.weixin.qq.com/s/n77GibL2corAtQHtVEAzfg)! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1REHvfQMKGnvz6p3Fd23HhSO4c8j5WPGZV0bKYLwnHyQ/edit?usp=sharing).
 - [2025/03] We hosted [the East Coast vLLM Meetup](https://lu.ma/7mu4k4xx)! Please find the meetup slides [here](https://docs.google.com/presentation/d/1NHiv8EUFF1NLd3fEYODm56nDmL26lEeXCaDgyDlTsRs/edit#slide=id.g31441846c39_0_0).
 - [2025/02] We hosted [the ninth vLLM meetup](https://lu.ma/h7g3kuj9) with Meta! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1jzC_PZVXrVNSFVCW-V4cFXb6pn7zZ2CyP_Flwo05aqg/edit?usp=sharing) and AMD [here](https://drive.google.com/file/d/1Zk5qEJIkTmlQ2eQcXQZlljAx3m9s7nwn/view?usp=sharing). The slides from Meta will not be posted.
 - [2025/01] We are excited to announce the alpha release of vLLM V1: A major architectural upgrade with 1.7x speedup! Clean code, optimized execution loop, zero-overhead prefix caching, enhanced multimodal support, and more. Please check out our blog post [here](https://blog.vllm.ai/2025/01/27/v1-alpha-release.html).
 - [2025/01] We hosted [the eighth vLLM meetup](https://lu.ma/zep56hui) with Google Cloud! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1epVkt4Zu8Jz_S5OhEHPc798emsYh2BwYfRuDDVEF7u4/edit?usp=sharing), and Google Cloud team [here](https://drive.google.com/file/d/1h24pHewANyRL11xy5dXUbvRC9F9Kkjix/view?usp=sharing).
 - [2024/12] vLLM joins [pytorch ecosystem](https://pytorch.org/blog/vllm-joins-pytorch)! Easy, Fast, and Cheap LLM Serving for Everyone!
+
+<details>
+<summary>Previous News</summary>
+
 - [2024/11] We hosted [the seventh vLLM meetup](https://lu.ma/h0qvrajz) with Snowflake! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1e3CxQBV3JsfGp30SwyvS3eM_tW-ghOhJ9PAJGK6KR54/edit?usp=sharing), and Snowflake team [here](https://docs.google.com/presentation/d/1qF3RkDAbOULwz9WK5TOltt2fE9t6uIc_hVNLFAaQX6A/edit?usp=sharing).
 - [2024/10] We have just created a developer slack ([slack.vllm.ai](https://slack.vllm.ai)) focusing on coordinating contributions and discussing features. Please feel free to join us there!
 - [2024/10] Ray Summit 2024 held a special track for vLLM! Please find the opening talk slides from the vLLM team [here](https://docs.google.com/presentation/d/1B_KQxpHBTRa_mDF-tR6i8rWdOU5QoTZNcEg2MKZxEHM/edit?usp=sharing). Learn more from the [talks](https://www.youtube.com/playlist?list=PLzTswPQNepXl6AQwifuwUImLPFRVpksjR) from other vLLM contributors and users!
@@ -34,8 +44,9 @@ Easy, fast, and cheap LLM serving for everyone
 - [2023/08] We would like to express our sincere gratitude to [Andreessen Horowitz](https://a16z.com/2023/08/30/supporting-the-open-source-ai-community/) (a16z) for providing a generous grant to support the open-source development and research of vLLM.
 - [2023/06] We officially released vLLM! FastChat-vLLM integration has powered [LMSYS Vicuna and Chatbot Arena](https://chat.lmsys.org) since mid-April. Check out our [blog post](https://vllm.ai).
 
----
+</details>
 
+---
 ## About
 
 vLLM is a fast and easy-to-use library for LLM inference and serving.
@@ -90,7 +101,7 @@ Visit our [documentation](https://docs.vllm.ai/en/latest/) to learn more.
 ## Contributing
 
 We welcome and value any contributions and collaborations.
-Please check out [CONTRIBUTING.md](./CONTRIBUTING.md) for how to get involved.
+Please check out [Contributing to vLLM](https://docs.vllm.ai/en/stable/contributing/overview.html) for how to get involved.
 
 ## Sponsors
 
@@ -113,6 +124,7 @@ Compute Resources:
 - Databricks
 - DeepInfra
 - Google Cloud
+- Intel
 - Lambda Lab
 - Nebius
 - Novita AI
@@ -143,10 +155,11 @@ If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs
 
 ## Contact Us
 
-- For technical questions and feature requests, please use GitHub issues or discussions.
-- For discussing with fellow users and coordinating contributions and development, please use Slack.
-- For security disclosures, please use GitHub's security advisory feature.
-- For collaborations and partnerships, please contact us at vllm-questions AT lists.berkeley.edu.
+- For technical questions and feature requests, please use GitHub [Issues](https://github.com/vllm-project/vllm/issues) or [Discussions](https://github.com/vllm-project/vllm/discussions)
+- For discussing with fellow users, please use the [vLLM Forum](https://discuss.vllm.ai)
+- coordinating contributions and development, please use [Slack](https://slack.vllm.ai)
+- For security disclosures, please use GitHub's [Security Advisories](https://github.com/vllm-project/vllm/security/advisories) feature
+- For collaborations and partnerships, please contact us at [vllm-questions@lists.berkeley.edu](mailto:vllm-questions@lists.berkeley.edu)
 
 ## Media Kit
 

benchmarks/README.md

@@ -41,29 +41,39 @@ become available.
       <td><code>synthetic</code></td>
     </tr>
     <tr>
-      <td><strong>HuggingFace</strong></td>
+      <td><strong>HuggingFace-VisionArena</strong></td>
       <td style="text-align: center;">✅</td>
-      <td style="text-align: center;">🟡</td>
-      <td>Specify your dataset path on HuggingFace</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>lmarena-ai/VisionArena-Chat</code></td>
     </tr>
     <tr>
-      <td><strong>VisionArena</strong></td>
+      <td><strong>HuggingFace-InstructCoder</strong></td>
       <td style="text-align: center;">✅</td>
       <td style="text-align: center;">✅</td>
-      <td><code>lmarena-ai/vision-arena-bench-v0.1</code> (a HuggingFace dataset)</td>
+      <td><code>likaixin/InstructCoder</code></td>
+    </tr>
+      <tr>
+      <td><strong>HuggingFace-AIMO</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>AI-MO/aimo-validation-aime</code> , <code>AI-MO/NuminaMath-1.5</code>, <code>AI-MO/NuminaMath-CoT</code></td>
+    </tr>
+    <tr>
+      <td><strong>HuggingFace-Other</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>lmms-lab/LLaVA-OneVision-Data</code>, <code>Aeala/ShareGPT_Vicuna_unfiltered</code></td>
     </tr>
   </tbody>
 </table>
 
 ✅: supported
 
+🟡: Partial support
+
 🚧: to be supported
 
-🟡: Partial support. Currently, HuggingFaceDataset only supports dataset formats
-similar to `lmms-lab/LLaVA-OneVision-Data`. If you need support for other dataset
-formats, please consider contributing.
-
-**Note**: VisionArena’s `dataset-name` should be set to `hf`
+**Note**: HuggingFace dataset's `dataset-name` should be set to `hf`
 
 ---
 ## Example - Online Benchmark
@@ -71,8 +81,7 @@ formats, please consider contributing.
 First start serving your model
 
 ```bash
-MODEL_NAME="NousResearch/Hermes-3-Llama-3.1-8B"
-vllm serve ${MODEL_NAME} --disable-log-requests
+vllm serve NousResearch/Hermes-3-Llama-3.1-8B --disable-log-requests
 ```
 
 Then run the benchmarking script
@@ -80,12 +89,13 @@ Then run the benchmarking script
 ```bash
 # download dataset
 # wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
-MODEL_NAME="NousResearch/Hermes-3-Llama-3.1-8B"
-NUM_PROMPTS=10
-BACKEND="vllm"
-DATASET_NAME="sharegpt"
-DATASET_PATH="<your data path>/ShareGPT_V3_unfiltered_cleaned_split.json"
-python3 vllm/benchmarks/benchmark_serving.py --backend ${BACKEND} --model ${MODEL_NAME} --endpoint /v1/completions --dataset-name ${DATASET_NAME} --dataset-path ${DATASET_PATH} --num-prompts ${NUM_PROMPTS}
+python3 vllm/benchmarks/benchmark_serving.py \
+  --backend vllm \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --endpoint /v1/completions \
+  --dataset-name sharegpt \
+  --dataset-path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \
+  --num-prompts 10
 ```
 
 If successful, you will see the following output
@@ -122,37 +132,87 @@ vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests
 ```
 
 ```bash
-MODEL_NAME="Qwen/Qwen2-VL-7B-Instruct"
-NUM_PROMPTS=10
-BACKEND="openai-chat"
-DATASET_NAME="hf"
-DATASET_PATH="lmarena-ai/vision-arena-bench-v0.1"
-DATASET_SPLIT='train'
-
 python3 vllm/benchmarks/benchmark_serving.py \
-  --backend "${BACKEND}" \
-  --model "${MODEL_NAME}" \
-  --endpoint "/v1/chat/completions" \
-  --dataset-name "${DATASET_NAME}" \
-  --dataset-path "${DATASET_PATH}" \
-  --hf-split "${DATASET_SPLIT}" \
-  --num-prompts "${NUM_PROMPTS}"
+  --backend openai-chat \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --endpoint /v1/chat/completions \
+  --dataset-name hf \
+  --dataset-path lmarena-ai/VisionArena-Chat \
+  --hf-split train \
+  --num-prompts 1000
+```
+
+### InstructCoder Benchmark with Speculative Decoding
+
+``` bash
+VLLM_USE_V1=1 vllm serve meta-llama/Meta-Llama-3-8B-Instruct \
+    --speculative-model "[ngram]" \
+    --ngram_prompt_lookup_min 2 \
+    --ngram-prompt-lookup-max 5 \
+    --num_speculative_tokens 5
+```
+
+``` bash
+python3 benchmarks/benchmark_serving.py \
+    --model meta-llama/Meta-Llama-3-8B-Instruct \
+    --dataset-name hf \
+    --dataset-path likaixin/InstructCoder \
+    --num-prompts 2048
+```
+
+### Other HuggingFaceDataset Examples
+
+```bash
+vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests
+```
+
+**`lmms-lab/LLaVA-OneVision-Data`**
+
+```bash
+python3 vllm/benchmarks/benchmark_serving.py \
+  --backend openai-chat \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --endpoint /v1/chat/completions \
+  --dataset-name hf \
+  --dataset-path lmms-lab/LLaVA-OneVision-Data \
+  --hf-split train \
+  --hf-subset "chart2text(cauldron)" \
+  --num-prompts 10
+```
+
+**`Aeala/ShareGPT_Vicuna_unfiltered`**
+
+```bash
+python3 vllm/benchmarks/benchmark_serving.py \
+  --backend openai-chat \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --endpoint /v1/chat/completions \
+  --dataset-name hf \
+  --dataset-path Aeala/ShareGPT_Vicuna_unfiltered \
+  --hf-split train \
+  --num-prompts 10
+```
+
+**`AI-MO/aimo-validation-aime`**
+
+``` bash
+python3 vllm/benchmarks/benchmark_serving.py \
+    --model Qwen/QwQ-32B \
+    --dataset-name hf \
+    --dataset-path AI-MO/aimo-validation-aime \
+    --num-prompts 10 \
+    --seed 42
 ```
 
 ---
 ## Example - Offline Throughput Benchmark
 
 ```bash
-MODEL_NAME="NousResearch/Hermes-3-Llama-3.1-8B"
-NUM_PROMPTS=10
-DATASET_NAME="sonnet"
-DATASET_PATH="vllm/benchmarks/sonnet.txt"
-
 python3 vllm/benchmarks/benchmark_throughput.py \
-  --model "${MODEL_NAME}" \
-  --dataset-name "${DATASET_NAME}" \
-  --dataset-path "${DATASET_PATH}" \
-  --num-prompts "${NUM_PROMPTS}"
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --dataset-name sonnet \
+  --dataset-path vllm/benchmarks/sonnet.txt \
+  --num-prompts 10
 ```
 
 If successful, you will see the following output
@@ -166,19 +226,13 @@ Total num output tokens:  1500
 ### VisionArena Benchmark for Vision Language Models
 
 ``` bash
-MODEL_NAME="Qwen/Qwen2-VL-7B-Instruct"
-NUM_PROMPTS=10
-DATASET_NAME="hf"
-DATASET_PATH="lmarena-ai/vision-arena-bench-v0.1"
-DATASET_SPLIT="train"
-
 python3 vllm/benchmarks/benchmark_throughput.py \
-  --model "${MODEL_NAME}" \
-  --backend "vllm-chat" \
-  --dataset-name "${DATASET_NAME}" \
-  --dataset-path "${DATASET_PATH}" \
-  --num-prompts "${NUM_PROMPTS}" \
-  --hf-split "${DATASET_SPLIT}"
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --backend vllm-chat \
+  --dataset-name hf \
+  --dataset-path lmarena-ai/VisionArena-Chat \
+  --num-prompts 1000 \
+  --hf-split train
 ```
 
 The `num prompt tokens` now includes image token counts
@@ -189,29 +243,83 @@ Total num prompt tokens:  14527
 Total num output tokens:  1280
 ```
 
+### InstructCoder Benchmark with Speculative Decoding
+
+``` bash
+VLLM_WORKER_MULTIPROC_METHOD=spawn \
+VLLM_USE_V1=1 \
+python3 vllm/benchmarks/benchmark_throughput.py \
+    --dataset-name=hf \
+    --dataset-path=likaixin/InstructCoder \
+    --model=meta-llama/Meta-Llama-3-8B-Instruct \
+    --input-len=1000 \
+    --output-len=100 \
+    --num-prompts=2048 \
+    --async-engine \
+    --speculative-model="[ngram]" \
+    --ngram_prompt_lookup_min=2 \
+    --ngram-prompt-lookup-max=5 \
+    --num_speculative_tokens=5
+```
+
+```
+Throughput: 104.77 requests/s, 23836.22 total tokens/s, 10477.10 output tokens/s
+Total num prompt tokens:  261136
+Total num output tokens:  204800
+```
+
+### Other HuggingFaceDataset Examples
+
+**`lmms-lab/LLaVA-OneVision-Data`**
+
+```bash
+python3 vllm/benchmarks/benchmark_throughput.py \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --backend vllm-chat \
+  --dataset-name hf \
+  --dataset-path lmms-lab/LLaVA-OneVision-Data \
+  --hf-split train \
+  --hf-subset "chart2text(cauldron)" \
+  --num-prompts 10
+```
+
+**`Aeala/ShareGPT_Vicuna_unfiltered`**
+
+```bash
+python3 vllm/benchmarks/benchmark_throughput.py \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --backend vllm-chat \
+  --dataset-name hf \
+  --dataset-path Aeala/ShareGPT_Vicuna_unfiltered \
+  --hf-split train \
+  --num-prompts 10
+```
+
+**`AI-MO/aimo-validation-aime`**
+
+```bash
+python3 benchmarks/benchmark_throughput.py \
+  --model Qwen/QwQ-32B \
+  --backend vllm \
+  --dataset-name hf \
+  --dataset-path AI-MO/aimo-validation-aime \
+  --hf-split train \
+  --num-prompts 10
+```
+
 ### Benchmark with LoRA Adapters
 
 ``` bash
 # download dataset
 # wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
-MODEL_NAME="meta-llama/Llama-2-7b-hf"
-BACKEND="vllm"
-DATASET_NAME="sharegpt"
-DATASET_PATH="<your data path>/ShareGPT_V3_unfiltered_cleaned_split.json"
-NUM_PROMPTS=10
-MAX_LORAS=2
-MAX_LORA_RANK=8
-ENABLE_LORA="--enable-lora"
-LORA_PATH="yard1/llama-2-7b-sql-lora-test"
-
 python3 vllm/benchmarks/benchmark_throughput.py \
-  --model "${MODEL_NAME}" \
-  --backend "${BACKEND}" \
-  --dataset_path "${DATASET_PATH}" \
-  --dataset_name "${DATASET_NAME}" \
-  --num-prompts "${NUM_PROMPTS}" \
-  --max-loras "${MAX_LORAS}" \
-  --max-lora-rank "${MAX_LORA_RANK}" \
-  ${ENABLE_LORA} \
-  --lora-path "${LORA_PATH}"
+  --model meta-llama/Llama-2-7b-hf \
+  --backend vllm \
+  --dataset_path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \
+  --dataset_name sharegpt \
+  --num-prompts 10 \
+  --max-loras 2 \
+  --max-lora-rank 8 \
+  --enable-lora \
+  --lora-path yard1/llama-2-7b-sql-lora-test
   ```

benchmarks/backend_request_func.py

@@ -63,7 +63,7 @@ async def async_request_tgi(
             "temperature": 0.01,  # TGI does not accept 0.0 temperature.
             "top_p": 0.99,  # TGI does not accept 1.0 top_p.
             "truncate": request_func_input.prompt_len,
-            # TGI does not accept ignore_eos flag.
+            "ignore_eos_token": request_func_input.ignore_eos,
         }
         payload = {
             "inputs": request_func_input.prompt,
@@ -71,6 +71,10 @@ async def async_request_tgi(
         }
         output = RequestFuncOutput()
         output.prompt_len = request_func_input.prompt_len
+        if request_func_input.ignore_eos:
+            output.output_tokens = request_func_input.output_len
+        else:
+            output.output_tokens = None
 
         ttft = 0.0
         st = time.perf_counter()
@@ -215,7 +219,15 @@ async def async_request_deepspeed_mii(
                 if response.status == 200:
                     parsed_resp = await response.json()
                     output.latency = time.perf_counter() - st
-                    output.generated_text = parsed_resp["text"][0]
+                    if "choices" in parsed_resp:
+                        output.generated_text = parsed_resp["choices"][0][
+                            "text"]
+                    elif "text" in parsed_resp:
+                        output.generated_text = parsed_resp["text"][0]
+                    else:
+                        output.error = ("Unexpected response format: "
+                                        "neither 'choices' nor 'text' found")
+                        output.success = False
                     output.success = True
                 else:
                     output.error = response.reason or ""

benchmarks/benchmark_dataset.py

@@ -17,12 +17,14 @@ SampleRequest instances, similar to the approach used in ShareGPT.
 import base64
 import io
 import json
+import logging
 import random
 from abc import ABC, abstractmethod
 from collections.abc import Mapping
 from dataclasses import dataclass
 from functools import cache
-from typing import Any, Optional, Union
+from io import BytesIO
+from typing import Any, Callable, Optional, Union
 
 import numpy as np
 import pandas as pd
@@ -35,6 +37,8 @@ from vllm.lora.utils import get_adapter_absolute_path
 from vllm.multimodal import MultiModalDataDict
 from vllm.transformers_utils.tokenizer import AnyTokenizer, get_lora_tokenizer
 
+logger = logging.getLogger(__name__)
+
 # -----------------------------------------------------------------------------
 # Data Classes
 # -----------------------------------------------------------------------------
@@ -61,9 +65,6 @@ class SampleRequest:
 class BenchmarkDataset(ABC):
     DEFAULT_SEED = 0
 
-    # num_requests has default 1000 in both the benchmark_serving.py and
-    # benchmark_throughput.py
-
     def __init__(
         self,
         dataset_path: Optional[str] = None,
@@ -90,8 +91,8 @@ class BenchmarkDataset(ABC):
             mm_content: Optional[MultiModalDataDict] = None) -> list[dict]:
         """
         Transform a prompt and optional multimodal content into a chat format.
-        This method is used for chat models that expect a specific 
-        conversation format.
+        This method is used for chat models that expect a specific conversation
+        format.
         """
         content = [{"text": prompt, "type": "text"}]
         if mm_content is not None:
@@ -101,10 +102,10 @@ class BenchmarkDataset(ABC):
     def load_data(self) -> None:
         """
         Load data from the dataset path into self.data.
-        
+
         This method must be overridden by subclasses since the method to load
         data will vary depending on the dataset format and source.
-        
+
         Raises:
             NotImplementedError: If a subclass does not implement this method.
         """
@@ -121,18 +122,18 @@ class BenchmarkDataset(ABC):
         """
         Optionally select a random LoRA request and return its associated
         tokenizer.
-        
+
         This method is used when LoRA parameters are provided.  It randomly
         selects a LoRA based on max_loras and retrieves a cached tokenizer for
         that LoRA if available. Otherwise, it returns the base tokenizer.
-        
+
         Args:
             tokenizer (PreTrainedTokenizerBase): The base tokenizer to use if no
             LoRA is selected.  max_loras (Optional[int]): The maximum number of
             LoRAs available. If None, LoRA is not used.  lora_path
             (Optional[str]): Path to the LoRA parameters on disk. If None, LoRA
             is not used.
-        
+
         Returns:
             tuple[Optional[LoRARequest], AnyTokenizer]: A tuple where the first
             element is a LoRARequest (or None if not applicable) and the second
@@ -160,21 +161,39 @@ class BenchmarkDataset(ABC):
                num_requests: int) -> list[SampleRequest]:
         """
         Abstract method to generate sample requests from the dataset.
-        
+
         Subclasses must override this method to implement dataset-specific logic
         for generating a list of SampleRequest objects.
-        
+
         Args:
             tokenizer (PreTrainedTokenizerBase): The tokenizer to be used
              for processing the dataset's text.
             num_requests (int): The number of sample requests to generate.
-        
+
         Returns:
             list[SampleRequest]: A list of sample requests generated from the
             dataset.
         """
         raise NotImplementedError("sample must be implemented in subclasses.")
 
+    def maybe_oversample_requests(self, requests: list[SampleRequest],
+                                  num_requests: int) -> None:
+        """
+        Oversamples the list of requests if its size is less than the desired
+        number.
+
+        Args:
+            requests (List[SampleRequest]): The current list of sampled
+            requests.  num_requests (int): The target number of requests.
+        """
+        if len(requests) < num_requests:
+            random.seed(self.random_seed)
+            additional = random.choices(requests,
+                                        k=num_requests - len(requests))
+            requests.extend(additional)
+            logger.info("Oversampled requests to reach %d total samples.",
+                        num_requests)
+
 
 # -----------------------------------------------------------------------------
 # Utility Functions and Global Caches
@@ -221,21 +240,24 @@ def process_image(image: Any) -> Mapping[str, Any]:
     """
     Process a single image input and return a multimedia content dictionary.
 
-    For a PIL.Image.Image input:
-      - Converts the image to RGB.
-      - Saves the image as a JPEG in-memory.
-      - Encodes the JPEG data as a base64 string.
-      - Returns a dictionary with the image as a base64 data URL.
+    Supports three input types:
 
-    For a string input:
-      - Treats the string as a URL or file path.
-      - Prepends "file://" if the string doesn't start with "http://" or
-        "file://".
-      - Returns a dictionary with the image URL.
+    1. Dictionary with raw image bytes: - Expects a dict with a 'bytes' key
+       containing raw image data.  - Loads the bytes as a PIL.Image.Image.
+
+    2. PIL.Image.Image input: - Converts the image to RGB.  - Saves the image as
+       a JPEG in memory.  - Encodes the JPEG data as a base64 string.  - Returns
+       a dictionary with the image as a base64 data URL.
+
+    3. String input: - Treats the string as a URL or local file path.  -
+       Prepends "file://" if the string doesn't start with "http://" or
+       "file://".  - Returns a dictionary with the image URL.
 
     Raises:
-      ValueError: If the input is neither a PIL.Image.Image nor a string.
+        ValueError: If the input is not a supported type.
     """
+    if isinstance(image, dict) and 'bytes' in image:
+        image = Image.open(BytesIO(image['bytes']))
     if isinstance(image, Image.Image):
         image = image.convert("RGB")
         with io.BytesIO() as image_data:
@@ -254,8 +276,8 @@ def process_image(image: Any) -> Mapping[str, Any]:
             ("http://", "file://")) else f"file://{image}")
         return {"type": "image_url", "image_url": {"url": image_url}}
 
-    raise ValueError(
-        f"Invalid image input {image}. Must be a PIL.Image.Image or str.")
+    raise ValueError(f"Invalid image input {image}. Must be a PIL.Image.Image"
+                     " or str or dictionary with raw image bytes.")
 
 
 # -----------------------------------------------------------------------------
@@ -276,15 +298,16 @@ class RandomDataset(BenchmarkDataset):
     ) -> None:
         super().__init__(**kwargs)
 
-    def sample(self,
-               tokenizer: PreTrainedTokenizerBase,
-               num_requests: int,
-               prefix_len: int = DEFAULT_PREFIX_LEN,
-               range_ratio: float = DEFAULT_RANGE_RATIO,
-               input_len: int = DEFAULT_INPUT_LEN,
-               output_len: int = DEFAULT_OUTPUT_LEN,
-               **kwargs) -> list[SampleRequest]:
-
+    def sample(
+        self,
+        tokenizer: PreTrainedTokenizerBase,
+        num_requests: int,
+        prefix_len: int = DEFAULT_PREFIX_LEN,
+        range_ratio: float = DEFAULT_RANGE_RATIO,
+        input_len: int = DEFAULT_INPUT_LEN,
+        output_len: int = DEFAULT_OUTPUT_LEN,
+        **kwargs,
+    ) -> list[SampleRequest]:
         vocab_size = tokenizer.vocab_size
 
         prefix_token_ids = (np.random.randint(
@@ -346,20 +369,24 @@ class ShareGPTDataset(BenchmarkDataset):
         random.seed(self.random_seed)
         random.shuffle(self.data)
 
-    def sample(self,
-               tokenizer: PreTrainedTokenizerBase,
-               num_requests: int,
-               lora_path: Optional[str] = None,
-               max_loras: Optional[int] = None,
-               output_len: Optional[int] = None,
-               enable_multimodal_chat: bool = False,
-               **kwargs) -> list:
+    def sample(
+        self,
+        tokenizer: PreTrainedTokenizerBase,
+        num_requests: int,
+        lora_path: Optional[str] = None,
+        max_loras: Optional[int] = None,
+        output_len: Optional[int] = None,
+        enable_multimodal_chat: bool = False,
+        **kwargs,
+    ) -> list:
         samples: list = []
         for entry in self.data:
             if len(samples) >= num_requests:
                 break
-            prompt, completion = entry["conversations"][0]["value"],\
-                entry["conversations"][1]["value"]
+            prompt, completion = (
+                entry["conversations"][0]["value"],
+                entry["conversations"][1]["value"],
+            )
 
             lora_request, tokenizer = self.get_random_lora_request(
                 tokenizer=tokenizer, max_loras=max_loras, lora_path=lora_path)
@@ -383,6 +410,7 @@ class ShareGPTDataset(BenchmarkDataset):
                     expected_output_len=new_output_len,
                     lora_request=lora_request,
                 ))
+        self.maybe_oversample_requests(samples, num_requests)
         return samples
 
 
@@ -415,19 +443,20 @@ class SonnetDataset(BenchmarkDataset):
         with open(self.dataset_path, encoding="utf-8") as f:
             self.data = f.readlines()
 
-    def sample(self,
-               tokenizer,
-               num_requests: int,
-               prefix_len: int = DEFAULT_PREFIX_LEN,
-               input_len: int = DEFAULT_INPUT_LEN,
-               output_len: int = DEFAULT_OUTPUT_LEN,
-               return_prompt_formatted: bool = False,
-               **kwargs) -> list:
+    def sample(
+        self,
+        tokenizer,
+        num_requests: int,
+        prefix_len: int = DEFAULT_PREFIX_LEN,
+        input_len: int = DEFAULT_INPUT_LEN,
+        output_len: int = DEFAULT_OUTPUT_LEN,
+        return_prompt_formatted: bool = False,
+        **kwargs,
+    ) -> list:
         # Calculate average token length for a poem line.
         tokenized_lines = [tokenizer(line).input_ids for line in self.data]
         avg_len = sum(len(tokens)
-                      for tokens in \
-                        tokenized_lines) / len(tokenized_lines)
+                      for tokens in tokenized_lines) / len(tokenized_lines)
 
         # Build the base prompt.
         base_prompt = "Pick as many lines as you can from these poem lines:\n"
@@ -506,12 +535,14 @@ class BurstGPTDataset(BenchmarkDataset):
         # Convert the dataframe to a list of lists.
         return data.values.tolist()
 
-    def sample(self,
-               tokenizer: PreTrainedTokenizerBase,
-               num_requests: int,
-               max_loras: Optional[int] = None,
-               lora_path: Optional[str] = None,
-               **kwargs) -> list[SampleRequest]:
+    def sample(
+        self,
+        tokenizer: PreTrainedTokenizerBase,
+        num_requests: int,
+        max_loras: Optional[int] = None,
+        lora_path: Optional[str] = None,
+        **kwargs,
+    ) -> list[SampleRequest]:
         samples = []
         data = self._sample_loaded_data(num_requests=num_requests)
         for i in range(num_requests):
@@ -535,49 +566,47 @@ class BurstGPTDataset(BenchmarkDataset):
 
 
 # -----------------------------------------------------------------------------
-# HuggingFace Dataset Implementation
+# HuggingFace Dataset Base Implementation
 # -----------------------------------------------------------------------------
-
-
 class HuggingFaceDataset(BenchmarkDataset):
-    """
-    Dataset class for processing a HuggingFace dataset with conversation data
-    and optional images.
-    """
-    DEFAULT_NUM_REQUESTS = 1000
+    """Base class for datasets hosted on HuggingFace."""
+
+    SUPPORTED_DATASET_PATHS: Union[set[str], dict[str, Callable]] = set()
 
     def __init__(
         self,
+        dataset_path: str,
         dataset_split: str,
         dataset_subset: Optional[str] = None,
         **kwargs,
     ) -> None:
-        super().__init__(**kwargs)
+        super().__init__(dataset_path=dataset_path, **kwargs)
+
         self.dataset_split = dataset_split
         self.dataset_subset = dataset_subset
-
         self.load_data()
 
     def load_data(self) -> None:
-        if not self.dataset_path:
-            raise ValueError("dataset_path must be provided for loading data.")
-
+        """Load data from HuggingFace datasets."""
         self.data = load_dataset(
             self.dataset_path,
             name=self.dataset_subset,
             split=self.dataset_split,
             streaming=True,
         )
-        if self.data.features is None or "conversations" \
-            not in self.data.features:
-            raise ValueError(
-                "HuggingFaceDataset currently only supports datasets with "
-                "a 'conversations' column like lmms-lab/LLaVA-OneVision-Data. "
-                "Please consider contributing if you would like to add "
-                "support for additional dataset formats.")
-        # Shuffle and filter examples with at least 2 conversations.
-        self.data = self.data.shuffle(seed=self.random_seed).filter(
-            lambda x: len(x["conversations"]) >= 2)
+        self.data = self.data.shuffle(seed=self.random_seed)
+
+
+# -----------------------------------------------------------------------------
+# Conversation Dataset Implementation
+# -----------------------------------------------------------------------------
+
+
+class ConversationDataset(HuggingFaceDataset):
+    """Dataset for conversation data with multimodal support."""
+    SUPPORTED_DATASET_PATHS = {
+        'lmms-lab/LLaVA-OneVision-Data', 'Aeala/ShareGPT_Vicuna_unfiltered'
+    }
 
     def sample(self,
                tokenizer: PreTrainedTokenizerBase,
@@ -585,10 +614,13 @@ class HuggingFaceDataset(BenchmarkDataset):
                output_len: Optional[int] = None,
                enable_multimodal_chat: bool = False,
                **kwargs) -> list:
+        # Filter examples with at least 2 conversations
+        filtered_data = self.data.filter(
+            lambda x: len(x["conversations"]) >= 2)
         sampled_requests = []
         dynamic_output = output_len is None
 
-        for item in self.data:
+        for item in filtered_data:
             if len(sampled_requests) >= num_requests:
                 break
             conv = item["conversations"]
@@ -618,6 +650,7 @@ class HuggingFaceDataset(BenchmarkDataset):
                     expected_output_len=output_len,
                     multi_modal_data=mm_content,
                 ))
+        self.maybe_oversample_requests(sampled_requests, num_requests)
         return sampled_requests
 
 
@@ -632,44 +665,32 @@ class VisionArenaDataset(HuggingFaceDataset):
     """
 
     DEFAULT_OUTPUT_LEN = 128
-    DEFAULT_NUM_REQUESTS = 1000
-    VISION_ARENA_DATASET_PATH = "lmarena-ai/vision-arena-bench-v0.1"
+    SUPPORTED_DATASET_PATHS = {
+        "lmarena-ai/VisionArena-Chat":
+        lambda x: x["conversation"][0][0]["content"],
+        "lmarena-ai/vision-arena-bench-v0.1":
+        lambda x: x["turns"][0][0]["content"]
+    }
 
-    def __init__(
+    def sample(
         self,
+        tokenizer: PreTrainedTokenizerBase,
+        num_requests: int,
+        output_len: Optional[int] = None,
+        enable_multimodal_chat: bool = False,
         **kwargs,
-    ) -> None:
-        super().__init__(**kwargs)
-        if self.dataset_path != self.VISION_ARENA_DATASET_PATH:
-            raise ValueError(f"Only support Vision Arena dataset.\
-                    This data path {self.dataset_path} is not valid.")
-        if self.dataset_subset is None and self.dataset_split != "train":
-            raise ValueError("Dataset split must be 'train'.")
-
-        self.load_data()
-
-    def load_data(self) -> None:
-        dataset = load_dataset(
-            self.dataset_path,
-            name=self.dataset_subset,
-            split=self.dataset_split,
-            streaming=True,
-        )
-        self.data = dataset.shuffle(seed=self.random_seed)
-
-    def sample(self,
-               tokenizer: PreTrainedTokenizerBase,
-               num_requests: int,
-               output_len: Optional[int] = None,
-               enable_multimodal_chat: bool = False,
-               **kwargs) -> list:
+    ) -> list:
         output_len = (output_len
                       if output_len is not None else self.DEFAULT_OUTPUT_LEN)
         sampled_requests = []
         for item in self.data:
             if len(sampled_requests) >= num_requests:
                 break
-            prompt = item["turns"][0][0]["content"]
+            parser_fn = self.SUPPORTED_DATASET_PATHS.get(self.dataset_path)
+            if parser_fn is None:
+                raise ValueError(
+                    f"Unsupported dataset path: {self.dataset_path}")
+            prompt = parser_fn(item)
             mm_content = process_image(item["images"][0])
             prompt_len = len(tokenizer(prompt).input_ids)
             if enable_multimodal_chat:
@@ -685,4 +706,98 @@ class VisionArenaDataset(HuggingFaceDataset):
                     expected_output_len=output_len,
                     multi_modal_data=mm_content,
                 ))
+        self.maybe_oversample_requests(sampled_requests, num_requests)
+        return sampled_requests
+
+
+# -----------------------------------------------------------------------------
+# Instruct Coder Dataset Implementation
+# -----------------------------------------------------------------------------
+
+
+class InstructCoderDataset(HuggingFaceDataset):
+    """
+    InstructCoder Dataset.
+    https://huggingface.co/datasets/likaixin/InstructCoder
+
+    InstructCoder is the dataset designed for general code editing.  It consists
+    of 114,239 instruction-input-output triplets, and covers multiple distinct
+    code editing scenario.
+    """
+
+    DEFAULT_OUTPUT_LEN = 200  # this is the average default output length
+    SUPPORTED_DATASET_PATHS = {
+        "likaixin/InstructCoder",
+    }
+
+    def sample(self,
+               tokenizer: PreTrainedTokenizerBase,
+               num_requests: int,
+               output_len: Optional[int] = None,
+               enable_multimodal_chat: bool = False,
+               **kwargs) -> list:
+        output_len = (output_len
+                      if output_len is not None else self.DEFAULT_OUTPUT_LEN)
+        sampled_requests = []
+        for item in self.data:
+            if len(sampled_requests) >= num_requests:
+                break
+            prompt = f"{item['instruction']}:\n{item['input']}"
+            prompt_len = len(tokenizer(prompt).input_ids)
+            sampled_requests.append(
+                SampleRequest(
+                    prompt=prompt,
+                    prompt_len=prompt_len,
+                    expected_output_len=output_len,
+                ))
+        self.maybe_oversample_requests(sampled_requests, num_requests)
+        return sampled_requests
+
+
+# -----------------------------------------------------------------------------
+# AIMO Dataset Implementation
+# -----------------------------------------------------------------------------
+
+
+class AIMODataset(HuggingFaceDataset):
+    """
+    Dataset class for processing a AIMO dataset with reasoning questions.
+    """
+    SUPPORTED_DATASET_PATHS = {
+        "AI-MO/aimo-validation-aime", "AI-MO/NuminaMath-1.5",
+        "AI-MO/NuminaMath-CoT"
+    }
+
+    def sample(self,
+               tokenizer: PreTrainedTokenizerBase,
+               num_requests: int,
+               output_len: Optional[int] = None,
+               **kwargs) -> list:
+        sampled_requests = []
+        dynamic_output = output_len is None
+
+        for item in self.data:
+            if len(sampled_requests) >= num_requests:
+                break
+            prompt, completion = item['problem'], item["solution"]
+
+            prompt_ids = tokenizer(prompt).input_ids
+            completion_ids = tokenizer(completion).input_ids
+            prompt_len = len(prompt_ids)
+            completion_len = len(completion_ids)
+            output_len = completion_len if dynamic_output else output_len
+            assert isinstance(output_len, int) and output_len > 0
+            if dynamic_output and not is_valid_sequence(prompt_len,
+                                                        completion_len,
+                                                        max_prompt_len=2048,
+                                                        max_total_len=32000):
+                continue
+            sampled_requests.append(
+                SampleRequest(
+                    prompt=prompt,
+                    prompt_len=prompt_len,
+                    expected_output_len=output_len,
+                    multi_modal_data=None,
+                ))
+        self.maybe_oversample_requests(sampled_requests, num_requests)
         return sampled_requests

benchmarks/benchmark_serving.py

@@ -7,9 +7,6 @@ On the server side, run one of the following commands:
         --swap-space 16 \
         --disable-log-requests
 
-    (TGI backend)
-    ./launch_tgi_server.sh <your_model> <max_batch_total_tokens>
-
 On the client side, run:
     python benchmarks/benchmark_serving.py \
         --backend <backend> \
@@ -52,9 +49,11 @@ try:
 except ImportError:
     from argparse import ArgumentParser as FlexibleArgumentParser
 
-from benchmark_dataset import (BurstGPTDataset, HuggingFaceDataset,
-                               RandomDataset, SampleRequest, ShareGPTDataset,
-                               SonnetDataset, VisionArenaDataset)
+from benchmark_dataset import (AIMODataset, BurstGPTDataset,
+                               ConversationDataset, HuggingFaceDataset,
+                               InstructCoderDataset, RandomDataset,
+                               SampleRequest, ShareGPTDataset, SonnetDataset,
+                               VisionArenaDataset)
 from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
 
 MILLISECONDS_TO_SECONDS_CONVERSION = 1000
@@ -586,19 +585,39 @@ def main(args: argparse.Namespace):
                                             return_prompt_formatted=True)
 
     elif args.dataset_name == "hf":
-        # Choose between VisionArenaDataset
-        # and HuggingFaceDataset based on provided parameters.
-        dataset_class = (VisionArenaDataset if args.dataset_path
-                         == VisionArenaDataset.VISION_ARENA_DATASET_PATH
-                         and args.hf_subset is None else HuggingFaceDataset)
+        # all following datasets are implemented from the
+        # HuggingFaceDataset base class
+        if args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS:
+            dataset_class = VisionArenaDataset
+            args.hf_split = "train"
+            args.hf_subset = None
+        elif args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS:
+            dataset_class = InstructCoderDataset
+            args.hf_split = "train"
+        elif args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS:
+            dataset_class = ConversationDataset
+        elif args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS:
+            dataset_class = AIMODataset
+            args.hf_split = "train"
+        else:
+            supported_datasets = set([
+                dataset_name for cls in HuggingFaceDataset.__subclasses__()
+                for dataset_name in cls.SUPPORTED_DATASET_PATHS
+            ])
+            raise ValueError(
+                f"Unsupported dataset path: {args.dataset_path}. "
+                "Huggingface dataset only supports dataset_path"
+                f" from one of following: {supported_datasets}. "
+                "Please consider contributing if you would "
+                "like to add support for additional dataset formats.")
         input_requests = dataset_class(
             dataset_path=args.dataset_path,
             dataset_subset=args.hf_subset,
             dataset_split=args.hf_split,
+            random_seed=args.seed,
         ).sample(
             num_requests=args.num_prompts,
             tokenizer=tokenizer,
-            random_seed=args.seed,
             output_len=args.hf_output_len,
         )
 

benchmarks/benchmark_serving_structured_output.py

@@ -5,9 +5,6 @@ On the server side, run one of the following commands:
     (vLLM OpenAI API server)
     vllm serve <your_model> --disable-log-requests
 
-    (TGI backend)
-    ./launch_tgi_server.sh <your_model> <max_batch_total_tokens>
-
 On the client side, run:
     python benchmarks/benchmark_serving_structured_output.py \
         --backend <backend> \
@@ -732,8 +729,11 @@ def main(args: argparse.Namespace):
         api_url = f"http://{args.host}:{args.port}{args.endpoint}"
         base_url = f"http://{args.host}:{args.port}"
 
-    tokenizer = get_tokenizer(tokenizer_id,
-                              trust_remote_code=args.trust_remote_code)
+    tokenizer = get_tokenizer(
+        tokenizer_id,
+        trust_remote_code=args.trust_remote_code,
+        tokenizer_mode=args.tokenizer_mode,
+    )
 
     if args.dataset == 'grammar':
         args.structure_type = 'guided_grammar'
@@ -876,6 +876,13 @@ if __name__ == "__main__":
         help=
         "Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
     )
+    parser.add_argument(
+        "--tokenizer-mode",
+        type=str,
+        default="auto",
+        help=
+        "Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
+    )
     parser.add_argument(
         "--num-prompts",
         type=int,
@@ -989,11 +996,12 @@ if __name__ == "__main__":
                         type=float,
                         default=1.0,
                         help="Ratio of Structured Outputs requests")
-    parser.add_argument("--structured-output-backend",
-                        type=str,
-                        choices=["outlines", "lm-format-enforcer", "xgrammar"],
-                        default="xgrammar",
-                        help="Backend to use for structured outputs")
+    parser.add_argument(
+        "--structured-output-backend",
+        type=str,
+        choices=["outlines", "lm-format-enforcer", "xgrammar", "guidance"],
+        default="xgrammar",
+        help="Backend to use for structured outputs")
 
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_throughput.py

@@ -11,7 +11,8 @@ from typing import Any, Optional, Union
 
 import torch
 import uvloop
-from benchmark_dataset import (BurstGPTDataset, HuggingFaceDataset,
+from benchmark_dataset import (AIMODataset, BurstGPTDataset,
+                               ConversationDataset, InstructCoderDataset,
                                RandomDataset, SampleRequest, ShareGPTDataset,
                                SonnetDataset, VisionArenaDataset)
 from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
@@ -300,6 +301,7 @@ def get_requests(args, tokenizer):
         "input_len": args.input_len,
         "output_len": args.output_len,
     }
+
     if args.dataset_path is None or args.dataset_name == "random":
         sample_kwargs["range_ratio"] = args.random_range_ratio
         sample_kwargs["prefix_len"] = args.prefix_len
@@ -317,18 +319,23 @@ def get_requests(args, tokenizer):
     elif args.dataset_name == "burstgpt":
         dataset_cls = BurstGPTDataset
     elif args.dataset_name == "hf":
-        if args.backend != "vllm-chat":
-            raise ValueError(
-                "hf datasets only are supported by vllm-chat backend")
-        # Choose between VisionArenaDataset and HuggingFaceDataset based on
-        # provided parameters.
-        dataset_cls = (VisionArenaDataset if args.dataset_path
-                       == VisionArenaDataset.VISION_ARENA_DATASET_PATH
-                       and args.hf_subset is None else HuggingFaceDataset)
-        common_kwargs['dataset_subset'] = args.hf_subset
-        common_kwargs['dataset_split'] = args.hf_split
-        sample_kwargs["enable_multimodal_chat"] = True
-
+        if args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS:
+            dataset_cls = VisionArenaDataset
+            common_kwargs['dataset_subset'] = None
+            common_kwargs['dataset_split'] = "train"
+            sample_kwargs["enable_multimodal_chat"] = True
+        elif args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS:
+            dataset_cls = InstructCoderDataset
+            common_kwargs['dataset_split'] = "train"
+        elif args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS:
+            dataset_cls = ConversationDataset
+            common_kwargs['dataset_subset'] = args.hf_subset
+            common_kwargs['dataset_split'] = args.hf_split
+            sample_kwargs["enable_multimodal_chat"] = True
+        elif args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS:
+            dataset_cls = AIMODataset
+            common_kwargs['dataset_subset'] = None
+            common_kwargs['dataset_split'] = "train"
     else:
         raise ValueError(f"Unknown dataset name: {args.dataset_name}")
     # Remove None values
@@ -462,9 +469,17 @@ def validate_args(args):
         warnings.warn("--hf-subset and --hf-split will be ignored \
                 since --dataset-name is not 'hf'.",
                       stacklevel=2)
-    elif args.dataset_name == "hf" and args.backend != "vllm-chat":
-        raise ValueError(
-            "When --dataset-name is 'hf', backend must be 'vllm-chat'")
+    elif args.dataset_name == "hf":
+        if args.dataset_path in (
+                VisionArenaDataset.SUPPORTED_DATASET_PATHS.keys()
+                | ConversationDataset.SUPPORTED_DATASET_PATHS):
+            assert args.backend == "vllm-chat", f"{args.dataset_path} needs to use vllm-chat as the backend."  #noqa: E501
+        elif args.dataset_path in (InstructCoderDataset.SUPPORTED_DATASET_PATHS
+                                   | AIMODataset.SUPPORTED_DATASET_PATHS):
+            assert args.backend == "vllm", f"{args.dataset_path} needs to use vllm as the backend."  #noqa: E501
+        else:
+            raise ValueError(
+                f"{args.dataset_path} is not supported by hf dataset.")
 
     # --random-range-ratio: only used when dataset_name is 'random'
     if args.dataset_name != 'random' and args.random_range_ratio is not None:

benchmarks/kernels/benchmark_grouped_gemm_cutlass.py

@@ -0,0 +1,340 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import torch
+import torch.utils.benchmark as benchmark
+from benchmark_shapes import WEIGHT_SHAPES_MOE
+
+from vllm import _custom_ops as ops
+from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config
+from vllm.model_executor.layers.fused_moe.fused_moe import (cutlass_moe_fp8,
+                                                            fused_experts,
+                                                            fused_topk)
+from vllm.utils import FlexibleArgumentParser
+
+DEFAULT_MODELS = [
+    "nm-testing/Mixtral-8x7B-Instruct-v0.1", "nm-testing/deepseekv2-lite",
+    "ibm-granite/granite-3.0-1b-a400m", "ibm-granite/granite-3.0-3b-a800m"
+]
+DEFAULT_BATCH_SIZES = [1, 4, 8, 16, 32, 64, 128, 256, 512]
+DEFAULT_TP_SIZES = [1]
+
+PER_ACT_TOKEN_OPTS = [False]
+PER_OUT_CH_OPTS = [False]
+
+
+def to_fp8(tensor: torch.Tensor):
+    finfo = torch.finfo(torch.float8_e4m3fn)
+    return torch.round(tensor.clamp(
+        min=finfo.min, max=finfo.max)).to(dtype=torch.float8_e4m3fn)
+
+
+def bench_run(results: list[benchmark.Measurement], model: str,
+              num_experts: int, topk: int, per_act_token: bool,
+              per_out_ch: bool, mkn: tuple[int, int, int]):
+    label = "Quant Matmul"
+
+    sub_label = (
+        "{}, num_experts={}, topk={}, per_act_token={} per_out_ch={}, "
+        "MKN=({})".format(model, num_experts, topk, per_act_token, per_out_ch,
+                          mkn))
+
+    print(f"Testing: {sub_label}")
+
+    (m, k, n) = mkn
+
+    dtype = torch.half
+
+    a = torch.randn((m, k), device="cuda", dtype=dtype) / 10
+    w1 = torch.randn((num_experts, 2 * n, k), device="cuda", dtype=dtype) / 10
+    w2 = torch.randn((num_experts, k, n), device="cuda", dtype=dtype) / 10
+
+    _, a_scale = ops.scaled_fp8_quant(a)
+
+    w1_q = torch.empty((num_experts, 2 * n, k),
+                       device="cuda",
+                       dtype=torch.float8_e4m3fn)
+    w2_q = torch.empty((num_experts, k, n),
+                       device="cuda",
+                       dtype=torch.float8_e4m3fn)
+    w1_scale = torch.empty((num_experts, 1, 1),
+                           device="cuda",
+                           dtype=torch.float32)
+    w2_scale = torch.empty((num_experts, 1, 1),
+                           device="cuda",
+                           dtype=torch.float32)
+
+    ab_strides1 = torch.full((num_experts, ),
+                             k,
+                             device="cuda",
+                             dtype=torch.int64)
+    c_strides1 = torch.full((num_experts, ),
+                            2 * n,
+                            device="cuda",
+                            dtype=torch.int64)
+    ab_strides2 = torch.full((num_experts, ),
+                             n,
+                             device="cuda",
+                             dtype=torch.int64)
+    c_strides2 = torch.full((num_experts, ),
+                            k,
+                            device="cuda",
+                            dtype=torch.int64)
+
+    for expert in range(num_experts):
+        w1_q[expert], w1_scale[expert] = ops.scaled_fp8_quant(w1[expert])
+        w2_q[expert], w2_scale[expert] = ops.scaled_fp8_quant(w2[expert])
+    w1_q_notransp = w1_q.clone()
+    w2_q_notransp = w2_q.clone()
+    w1_q = w1_q.transpose(1, 2)
+    w2_q = w2_q.transpose(1, 2)
+
+    score = torch.randn((m, num_experts), device="cuda", dtype=dtype)
+
+    topk_weights, topk_ids = fused_topk(a, score, topk, renormalize=False)
+
+    def run_triton_moe(a: torch.Tensor, w1: torch.Tensor, w2: torch.Tensor,
+                       topk_weights: torch.Tensor, topk_ids: torch.Tensor,
+                       w1_scale: torch.Tensor, w2_scale: torch.Tensor,
+                       a_scale: torch.Tensor, num_repeats: int):
+        for _ in range(num_repeats):
+            fused_experts(a,
+                          w1,
+                          w2,
+                          topk_weights,
+                          topk_ids,
+                          use_fp8_w8a8=True,
+                          w1_scale=w1_scale,
+                          w2_scale=w2_scale,
+                          a1_scale=a_scale)
+
+    def run_cutlass_moe(a: torch.Tensor, a_scale: torch.Tensor,
+                        w1: torch.Tensor, w2: torch.Tensor,
+                        w1_scale: torch.Tensor, w2_scale: torch.Tensor,
+                        topk_weights: torch.Tensor, topk_ids: torch.Tensor,
+                        ab_strides1: torch.Tensor, c_strides1: torch.Tensor,
+                        ab_strides2: torch.Tensor, c_strides2: torch.Tensor,
+                        num_repeats: int):
+        for _ in range(num_repeats):
+            cutlass_moe_fp8(a,
+                            w1,
+                            w2,
+                            w1_scale,
+                            w2_scale,
+                            topk_weights,
+                            topk_ids,
+                            ab_strides1,
+                            c_strides1,
+                            ab_strides2,
+                            c_strides2,
+                            a1_scale=a_scale)
+
+    def run_cutlass_from_graph(
+            a: torch.Tensor, a_scale: torch.Tensor, w1_q: torch.Tensor,
+            w2_q: torch.Tensor, w1_scale: torch.Tensor, w2_scale: torch.Tensor,
+            topk_weights: torch.Tensor, topk_ids: torch.Tensor,
+            ab_strides1: torch.Tensor, c_strides1: torch.Tensor,
+            ab_strides2: torch.Tensor, c_strides2: torch.Tensor):
+        with set_current_vllm_config(
+                VllmConfig(parallel_config=ParallelConfig(
+                    pipeline_parallel_size=1))):
+            return cutlass_moe_fp8(a,
+                                   w1_q,
+                                   w2_q,
+                                   w1_scale,
+                                   w2_scale,
+                                   topk_weights,
+                                   topk_ids,
+                                   ab_strides1,
+                                   c_strides1,
+                                   ab_strides2,
+                                   c_strides2,
+                                   a1_scale=a_scale)
+
+    def run_triton_from_graph(a: torch.Tensor, w1: torch.Tensor,
+                              w2: torch.Tensor, topk_weights: torch.Tensor,
+                              topk_ids: torch.Tensor, w1_scale: torch.Tensor,
+                              w2_scale: torch.Tensor, a_scale: torch.Tensor):
+        with set_current_vllm_config(
+                VllmConfig(parallel_config=ParallelConfig(
+                    pipeline_parallel_size=1))):
+            return fused_experts(a,
+                                 w1,
+                                 w2,
+                                 topk_weights,
+                                 topk_ids,
+                                 use_fp8_w8a8=True,
+                                 w1_scale=w1_scale,
+                                 w2_scale=w2_scale,
+                                 a1_scale=a_scale)
+
+    def replay_graph(graph, num_repeats):
+        for _ in range(num_repeats):
+            graph.replay()
+        torch.cuda.synchronize()
+
+    cutlass_stream = torch.cuda.Stream()
+    cutlass_graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(cutlass_graph, stream=cutlass_stream):
+        run_cutlass_from_graph(a, a_scale, w1_q, w2_q, w1_scale, w2_scale,
+                               topk_weights, topk_ids, ab_strides1, c_strides1,
+                               ab_strides2, c_strides2)
+    torch.cuda.synchronize()
+
+    triton_stream = torch.cuda.Stream()
+    triton_graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(triton_graph, stream=triton_stream):
+        run_triton_from_graph(a, w1_q_notransp, w2_q_notransp, topk_weights,
+                              topk_ids, w1_scale, w2_scale, a_scale)
+    torch.cuda.synchronize()
+
+    min_run_time = 5
+    num_warmup = 5
+    num_runs = 25
+
+    globals = {
+        # Baseline params
+        "w1": w1,
+        "w2": w2,
+        "score": score,
+        "topk": topk,
+        "w1_q_notransp": w1_q_notransp,
+        "w2_q_notransp": w2_q_notransp,
+        # Cutlass params
+        "a_scale": a_scale,
+        "w1_q": w1_q,
+        "w2_q": w2_q,
+        "w1_scale": w1_scale,
+        "w2_scale": w2_scale,
+        "ab_strides1": ab_strides1,
+        "c_strides1": c_strides1,
+        "ab_strides2": ab_strides2,
+        "c_strides2": c_strides2,
+        # cuda graph params
+        "cutlass_graph": cutlass_graph,
+        "triton_graph": triton_graph,
+        # Gen params
+        "a": a,
+        "topk_weights": topk_weights,
+        "topk_ids": topk_ids,
+        "num_runs": num_runs,
+        # Kernels
+        "run_triton_moe": run_triton_moe,
+        "run_cutlass_moe": run_cutlass_moe,
+        "replay_graph": replay_graph,
+    }
+
+    # Warmup
+    run_triton_moe(a, w1_q_notransp, w2_q_notransp, topk_weights, topk_ids,
+                   w1_scale, w2_scale, a_scale, num_warmup)
+
+    results.append(
+        benchmark.Timer(
+            stmt=
+            "run_triton_moe(a, w1_q_notransp, w2_q_notransp, topk_weights, topk_ids, w1_scale, w2_scale, a_scale, num_runs)",  # noqa: E501
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="triton_moe",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+    # Warmup
+    replay_graph(triton_graph, num_warmup)
+
+    results.append(
+        benchmark.Timer(
+            stmt="replay_graph(triton_graph, num_runs)",
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="triton_moe_cuda_graphs",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+    # Warmup
+    run_cutlass_moe(a, a_scale, w1_q, w2_q, w1_scale, w2_scale, topk_weights,
+                    topk_ids, ab_strides1, c_strides1, ab_strides2, c_strides2,
+                    num_warmup)
+
+    results.append(
+        benchmark.Timer(
+            stmt=
+            "run_cutlass_moe(a, a_scale, w1_q, w2_q, w1_scale, w2_scale, topk_weights, topk_ids, ab_strides1, c_strides1, ab_strides2, c_strides2, num_runs)",  # noqa: E501
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="grouped_gemm_moe",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+    # Warmup
+    replay_graph(cutlass_graph, num_warmup)
+
+    results.append(
+        benchmark.Timer(
+            stmt="replay_graph(cutlass_graph, num_runs)",
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="grouped_gemm_moe_cuda_graphs",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+
+def main(args):
+    print("Benchmarking models:")
+    for i, model in enumerate(args.models):
+        print(f"[{i}]  {model}")
+
+    results: list[benchmark.Measurement] = []
+
+    for model in args.models:
+        for tp in args.tp_sizes:
+            for layer in WEIGHT_SHAPES_MOE[model]:
+                num_experts = layer[0]
+                topk = layer[1]
+                size_k = layer[2]
+                size_n = layer[3] // tp
+
+                if len(args.limit_k) > 0 and size_k not in args.limit_k:
+                    continue
+
+                if len(args.limit_n) > 0 and size_n not in args.limit_n:
+                    continue
+
+                for per_act_token in PER_ACT_TOKEN_OPTS:
+                    for per_out_ch in PER_OUT_CH_OPTS:
+                        for size_m in DEFAULT_BATCH_SIZES:
+                            mkn = (size_m, size_k, size_n)
+                            bench_run(results, model, num_experts, topk,
+                                      per_act_token, per_out_ch, mkn)
+
+    compare = benchmark.Compare(results)
+    compare.print()
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(
+        description="Benchmark Marlin across specified models/shapes/batches")
+    parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=DEFAULT_MODELS,
+        choices=WEIGHT_SHAPES_MOE.keys(),
+    )
+    parser.add_argument("--tp-sizes",
+                        nargs="+",
+                        type=int,
+                        default=DEFAULT_TP_SIZES)
+    parser.add_argument("--batch-sizes",
+                        nargs="+",
+                        type=int,
+                        default=DEFAULT_BATCH_SIZES)
+    parser.add_argument("--limit-k", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-n", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-num-groups", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-per-act-token",
+                        nargs="+",
+                        type=int,
+                        default=[])
+    parser.add_argument("--limit-per-out-ch", nargs="+", type=int, default=[])
+
+    args = parser.parse_args()
+    main(args)

benchmarks/kernels/benchmark_lora.py

@@ -17,13 +17,8 @@ from torch.utils.benchmark import Measurement as TMeasurement
 from utils import ArgPool, Bench, CudaGraphBenchParams
 from weight_shapes import WEIGHT_SHAPES
 
-from vllm.lora.ops.triton_ops.bgmv_expand import bgmv_expand
-from vllm.lora.ops.triton_ops.bgmv_expand_slice import bgmv_expand_slice
-from vllm.lora.ops.triton_ops.bgmv_shrink import bgmv_shrink
-from vllm.lora.ops.triton_ops.sgmv_expand import sgmv_expand
-from vllm.lora.ops.triton_ops.sgmv_shrink import sgmv_shrink
+from vllm.lora.ops.triton_ops import LoRAKernelMeta, lora_expand, lora_shrink
 from vllm.lora.ops.triton_ops.utils import _LORA_A_PTR_DICT, _LORA_B_PTR_DICT
-from vllm.lora.ops.triton_ops.v1 import V1KernelMeta, v1_expand, v1_shrink
 from vllm.utils import FlexibleArgumentParser
 
 DEFAULT_MODELS = list(WEIGHT_SHAPES.keys())
@@ -167,69 +162,25 @@ class OpType(Enum):
     """
     LoRA Ops to benchmark and its properties.
     """
-    SGMV_SHRINK = auto()
-    BGMV_SHRINK = auto()
-    SGMV_EXPAND = auto()
-    BGMV_EXPAND = auto()
-    BGMV_EXPAND_SLICE = auto()
-    V1_SHRINK = auto()
-    V1_EXPAND = auto()
+    LORA_SHRINK = auto()
+    LORA_EXPAND = auto()
 
     @staticmethod
     def from_str(s: str) -> "OpType":
-        if s.lower() == 'sgmv_shrink':
-            return OpType.SGMV_SHRINK
-        if s.lower() == 'sgmv_expand':
-            return OpType.SGMV_EXPAND
-        if s.lower() == 'bgmv_shrink':
-            return OpType.BGMV_SHRINK
-        if s.lower() == 'bgmv_expand':
-            return OpType.BGMV_EXPAND
-        if s.lower() == "bgmv_expand_slice":
-            return OpType.BGMV_EXPAND_SLICE
-        if s.lower() == "v1_shrink":
-            return OpType.V1_SHRINK
-        if s.lower() == "v1_expand":
-            return OpType.V1_EXPAND
+        if s.lower() == "lora_shrink":
+            return OpType.LORA_SHRINK
+        if s.lower() == "lora_expand":
+            return OpType.LORA_EXPAND
         raise ValueError(f"Unrecognized str {s} to convert to OpType")
 
     def is_shrink_fn(self) -> bool:
-        return self in [
-            OpType.SGMV_SHRINK, OpType.BGMV_SHRINK, OpType.V1_SHRINK
-        ]
+        return self in [OpType.LORA_SHRINK]
 
     def is_expand_fn(self) -> bool:
-        return self in [
-            OpType.SGMV_EXPAND, OpType.BGMV_EXPAND, OpType.V1_EXPAND
-        ]
-
-    def is_prefill_op(self) -> bool:
-        return self in [
-            OpType.SGMV_SHRINK, OpType.SGMV_EXPAND, OpType.V1_SHRINK,
-            OpType.V1_EXPAND
-        ]
-
-    def is_decode_op(self) -> bool:
-        return self in [
-            OpType.BGMV_SHRINK, OpType.BGMV_EXPAND, OpType.BGMV_EXPAND_SLICE,
-            OpType.V1_SHRINK, OpType.V1_EXPAND
-        ]
-
-    def is_expand_slice_fn(self) -> bool:
-        return self in [OpType.BGMV_EXPAND_SLICE]
+        return self in [OpType.LORA_EXPAND]
 
     def num_slices(self) -> list[int]:
-        if self in [
-                OpType.SGMV_EXPAND, OpType.SGMV_SHRINK, OpType.V1_SHRINK,
-                OpType.V1_EXPAND
-        ]:
-            # SGMV kernels and v1 kernels supports slices
-            return [1, 2, 3]
-        if self in [OpType.BGMV_SHRINK, OpType.BGMV_EXPAND]:
-            return [1]
-        if self in [OpType.BGMV_EXPAND_SLICE]:
-            return [2, 3]
-        raise ValueError(f"Unrecognized OpType {self}")
+        return [1, 2, 3]
 
     def mkn(self, batch_size: int, seq_length: int, hidden_size: int,
             lora_rank: int) -> tuple[int, int, int]:
@@ -239,7 +190,7 @@ class OpType(Enum):
             k = hidden_size
             n = lora_rank
         else:
-            assert self.is_expand_fn() or self.is_expand_slice_fn()
+            assert self.is_expand_fn()
             m = num_tokens
             k = lora_rank
             n = hidden_size
@@ -254,7 +205,7 @@ class OpType(Enum):
         if self.is_shrink_fn():
             return op_dtype, op_dtype, torch.float32
         else:
-            assert self.is_expand_fn() or self.is_expand_slice_fn()
+            assert self.is_expand_fn()
             return torch.float32, op_dtype, op_dtype
 
     def matmul_shapes(
@@ -268,43 +219,19 @@ class OpType(Enum):
         m, k, n = self.mkn(batch_size, seq_length, hidden_size, lora_rank)
 
         b_shape = (num_loras, n, k)  # col-major
-        if self in [OpType.SGMV_SHRINK, OpType.V1_SHRINK]:
-            # SGMV shrink and V1 shrink kernels support num_slices inherently
-            # in the kernel.
+        if self in [OpType.LORA_SHRINK]:
+            # LoRA shrink kernels support num_slices inherently in the kernel.
             return ((m, k), b_shape, (num_slices, m, n))
-        if self in [OpType.SGMV_EXPAND, OpType.V1_EXPAND]:
-            # SGMV expand and V1 expand kernels support num_slices inherently
-            # in the kernel
+        if self in [OpType.LORA_EXPAND]:
+            # LoRA expand kernels support num_slices inherently in the kernel
             return ((num_slices, m, k), b_shape, (m, n * num_slices))
-        if self == OpType.BGMV_SHRINK:
-            return ((m, k), b_shape, (m, n))
-        if self == OpType.BGMV_EXPAND:
-            return ((m, k), b_shape, (m, n))
-        if self == OpType.BGMV_EXPAND_SLICE:
-            return ((num_slices, m, k), b_shape, (m, n * num_slices))
-
         raise ValueError(f"Unrecognized op_type {self}")
 
     def bench_fn(self) -> Callable:
-
-        def emulate_bgmv_expand_slice(kwargs_list: list[dict[str, Any]]):
-            for x in kwargs_list:
-                bgmv_expand_slice(**x)
-
-        if self == OpType.SGMV_SHRINK:
-            return sgmv_shrink
-        if self == OpType.SGMV_EXPAND:
-            return sgmv_expand
-        if self == OpType.BGMV_SHRINK:
-            return bgmv_shrink
-        if self == OpType.BGMV_EXPAND:
-            return bgmv_expand
-        if self == OpType.BGMV_EXPAND_SLICE:
-            return emulate_bgmv_expand_slice
-        if self == OpType.V1_SHRINK:
-            return v1_shrink
-        if self == OpType.V1_EXPAND:
-            return v1_expand
+        if self == OpType.LORA_SHRINK:
+            return lora_shrink
+        if self == OpType.LORA_EXPAND:
+            return lora_expand
 
         raise ValueError(f"Unrecognized optype {self}")
 
@@ -318,34 +245,13 @@ class OpType(Enum):
         """
         w_dtype = lora_weights[0].dtype
         num_slices = len(lora_weights)
-        if self in [OpType.SGMV_SHRINK, OpType.V1_SHRINK]:
+        if self in [OpType.LORA_SHRINK]:
             for slice_idx in range(num_slices):
                 ref_group_gemm(ref_out=output[slice_idx, :],
                                input=input,
                                lora_weights=lora_weights[slice_idx],
                                **kwargs)
-        elif self in [OpType.SGMV_EXPAND, OpType.V1_EXPAND]:
-            hidden_size = lora_weights[0].shape[1]
-            for slice_idx in range(num_slices):
-                slice_offset = slice_idx * hidden_size
-                ref_group_gemm(
-                    ref_out=output[:, slice_offset:slice_offset + hidden_size],
-                    input=input[slice_idx].clone().to(dtype=w_dtype),
-                    lora_weights=lora_weights[slice_idx],
-                    **kwargs)
-        elif self == OpType.BGMV_SHRINK:
-            assert num_slices == 1
-            ref_group_gemm(ref_out=output,
-                           input=input,
-                           lora_weights=lora_weights[0],
-                           **kwargs)
-        elif self == OpType.BGMV_EXPAND:
-            assert num_slices == 1
-            ref_group_gemm(ref_out=output,
-                           input=input.clone().to(dtype=w_dtype),
-                           lora_weights=lora_weights[0],
-                           **kwargs)
-        elif self == OpType.BGMV_EXPAND_SLICE:
+        elif self in [OpType.LORA_EXPAND]:
             hidden_size = lora_weights[0].shape[1]
             for slice_idx in range(num_slices):
                 slice_offset = slice_idx * hidden_size
@@ -411,13 +317,11 @@ class BenchmarkTensors:
     input: torch.Tensor
     lora_weights_lst: list[torch.Tensor]
     output: torch.Tensor
-    # metadata tensors
+    # LoRA kernel metadata
+    lora_kernel_meta: LoRAKernelMeta
+    # Metadata tensors used in testing correctness
     seq_lens: torch.Tensor
-    seq_start_loc: torch.Tensor
     prompt_lora_mapping: torch.Tensor
-    token_lora_mapping: torch.Tensor
-    # v1 kernel metadata
-    v1_kernel_meta: Optional[V1KernelMeta] = None
 
     def io_types(self) -> str:
         return (f"{dtype_to_str(self.input.dtype)}x"
@@ -444,35 +348,29 @@ class BenchmarkTensors:
         assert ctx.num_active_loras <= ctx.num_loras
         total_tokens = ctx.batch_size * ctx.seq_length
 
+        # Make metadata tensors involved in correctness testing.
         # Prepare seq lens tensor
         seq_len_tensor = torch.randint(ctx.seq_length, ctx.seq_length + 1,
                                        (ctx.batch_size, ))
-        # Prepare seq_start_loc tensor
-        seq_start_loc_tensor = torch.cumsum(torch.tensor(
-            [0] + seq_len_tensor[:-1].tolist(), dtype=torch.long),
-                                            dim=0)
         assert total_tokens == seq_len_tensor.sum()
         # Prepare prompt lora indices tensor
         prompt_lora_indices_tensor = make_prompt_lora_mapping(
             ctx.batch_size, ctx.num_active_loras, ctx.sort_by_lora_id, "cpu")
-        # Prepare token lora indices tensor
+
+        # Make LoRAKernelMeta
         token_lora_indices_tensor = make_token_lora_mapping(
             total_tokens, ctx.batch_size, prompt_lora_indices_tensor,
             seq_len_tensor, "cpu")
-
-        v1_kernel_meta = None
-        if op_type in [OpType.V1_SHRINK, OpType.V1_EXPAND]:
-            v1_kernel_meta = V1KernelMeta.make(
-                max_loras=ctx.num_loras,
-                max_num_tokens=token_lora_indices_tensor.size(0),
-                device="cpu")
-            v1_kernel_meta.prepare_tensors(
-                token_lora_mapping=token_lora_indices_tensor)
+        lora_kernel_meta = LoRAKernelMeta.make(
+            max_loras=ctx.num_loras,
+            max_num_tokens=token_lora_indices_tensor.size(0),
+            device="cpu")
+        lora_kernel_meta.prepare_tensors(
+            token_lora_mapping=token_lora_indices_tensor)
 
         return BenchmarkTensors(input_tensor, lora_weights, output_tensor,
-                                seq_len_tensor, seq_start_loc_tensor,
-                                prompt_lora_indices_tensor,
-                                token_lora_indices_tensor, v1_kernel_meta)
+                                lora_kernel_meta, seq_len_tensor,
+                                prompt_lora_indices_tensor)
 
     def sanity_check(self) -> None:
         """
@@ -482,9 +380,9 @@ class BenchmarkTensors:
         # check metadata tensors
         assert torch.sum(self.seq_lens) == num_tokens
         num_seqs = self.seq_lens.shape[0]
-        assert self.seq_start_loc.shape[0] == num_seqs
+        #assert self.seq_start_loc.shape[0] == num_seqs
         assert self.prompt_lora_mapping.shape[0] == num_seqs
-        assert self.token_lora_mapping.shape[0] == num_tokens
+        assert self.lora_kernel_meta.token_lora_mapping.shape[0] == num_tokens
 
     def to_device(self, device: str):
         """
@@ -499,220 +397,27 @@ class BenchmarkTensors:
         self.input = to_device(self.input)
         self.output = to_device(self.output)
         self.seq_lens = to_device(self.seq_lens)
-        self.seq_start_loc = to_device(self.seq_start_loc)
         self.prompt_lora_mapping = to_device(self.prompt_lora_mapping)
-        self.token_lora_mapping = to_device(self.token_lora_mapping)
         for i in range(len(self.lora_weights_lst)):
             self.lora_weights_lst[i] = to_device(self.lora_weights_lst[i])
 
-        # v1 meta
-        if self.v1_kernel_meta:
-            for field_name in V1KernelMeta.__dataclass_fields__:
-                field = getattr(self.v1_kernel_meta, field_name)
-                assert isinstance(field, torch.Tensor)
-                setattr(self.v1_kernel_meta, field_name, to_device(field))
+        # LoRA meta
+        for field_name in LoRAKernelMeta.__dataclass_fields__:
+            field = getattr(self.lora_kernel_meta, field_name)
+            assert isinstance(field, torch.Tensor)
+            setattr(self.lora_kernel_meta, field_name, to_device(field))
 
     def metadata(self) -> tuple[int, int, int]:
         """
         Return num_seqs, num_tokens and max_seq_len
         """
         num_seqs = self.seq_lens.shape[0]
-        num_tokens = self.token_lora_mapping.shape[0]
+        num_tokens = self.lora_kernel_meta.token_lora_mapping.shape[0]
         max_seq_len = torch.max(self.seq_lens).item()
         num_slices = len(self.lora_weights_lst)
         return num_seqs, num_tokens, max_seq_len, num_slices
 
-    def convert_to_sgmv_benchmark_tensors(self):
-        """
-        For sgmv punica kernels, when consecutive sequences have the
-        same LoRA ID, we just merge them together.
-        This happens in punica.py::compute_metadata
-        """
-
-        # Collapse seq_lens and seq_start_loc
-        _, seq_lens = torch.unique_consecutive(self.token_lora_mapping,
-                                               return_counts=True)
-        cum_result = torch.cumsum(seq_lens, dim=0)
-        seq_start_loc = torch.zeros_like(seq_lens)
-        seq_start_loc[1:].copy_(cum_result[:-1])
-
-        # Collapse prompt mapping
-        prompt_lora_mapping = torch.unique_consecutive(
-            self.prompt_lora_mapping)
-
-        assert torch.sum(seq_lens) == torch.sum(self.seq_lens), \
-         f"dont match - new {torch.sum(seq_lens)} vs {torch.sum(self.seq_lens)}"
-
-        self.prompt_lora_mapping = prompt_lora_mapping.to(
-            dtype=self.prompt_lora_mapping.dtype)
-        self.seq_lens = seq_lens.to(dtype=self.seq_lens.dtype)
-        self.seq_start_loc = seq_start_loc.to(dtype=self.seq_start_loc.dtype)
-
-    def as_sgmv_shrink_kwargs(self) -> dict[str, Any]:
-        self.convert_to_sgmv_benchmark_tensors()
-        self.sanity_check()
-        self.to_device(self.input.device)
-
-        num_seqs, num_tokens, max_seq_len, num_slices = self.metadata()
-
-        # Sanity check matrix shapes.
-        i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[
-            0].shape, self.output.shape
-        # Expected input shape [num_tokens, hidden_size]
-        assert len(i_shape) == 2
-        assert i_shape[0] == num_tokens
-        hidden_size = i_shape[1]
-        # Expected lora weight shape [num_loras, lora_rank, hidden_size]
-        assert len(lw_shape) == 3
-        assert lw_shape[2] == hidden_size
-        lora_rank = lw_shape[1]
-        # Expected output shape [num_slices, num_tokens, lora_rank]
-        assert len(o_shape) == 3
-        assert o_shape == (num_slices, num_tokens, lora_rank)
-
-        return {
-            'inputs': self.input,
-            'lora_a_weights': self.lora_weights_lst,
-            'output_tensor': self.output,
-            'b_seq_start_loc': self.seq_start_loc,
-            'seq_len_tensor': self.seq_lens,
-            'lora_indices_tensor': self.prompt_lora_mapping,
-            'batches': num_seqs,
-            'max_seq_length': max_seq_len,
-            'token_nums': num_tokens,
-            'scaling': 1.0,
-        }
-
-    def as_sgmv_expand_kwargs(self, add_inputs: bool) -> dict[str, Any]:
-
-        self.convert_to_sgmv_benchmark_tensors()
-        self.sanity_check()
-        self.to_device(self.input.device)
-
-        num_seqs, num_tokens, max_seq_len, num_slices = self.metadata()
-
-        # Sanity check matrix shapes.
-        i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[
-            0].shape, self.output.shape
-        # Expected input shape : [num_slices, num_tokens, lora_rank]
-        assert len(i_shape) == 3
-        assert i_shape[0] == num_slices
-        assert i_shape[1] == num_tokens
-        lora_rank = i_shape[2]
-        # Expected lora weight shape : [num_lora, hidden_size, lora_rank]
-        assert len(lw_shape) == 3
-        assert lw_shape[2] == lora_rank
-        hidden_size = lw_shape[1]
-        # Expected output shape : [num_tokens, hidden_size * num_slices]
-        assert len(o_shape) == 2
-        assert o_shape == (num_tokens, hidden_size * num_slices)
-
-        return {
-            'inputs': self.input,
-            'lora_b_weights': self.lora_weights_lst,
-            'output_tensor': self.output,
-            'b_seq_start_loc': self.seq_start_loc,
-            'seq_len_tensor': self.seq_lens,
-            'lora_indices_tensor': self.prompt_lora_mapping,
-            'batches': num_seqs,
-            'max_seq_length': max_seq_len,
-            'token_nums': num_tokens,
-            'offset_start': 0,
-            'add_inputs': add_inputs,
-        }
-
-    def as_bgmv_shrink_kwargs(self) -> dict[str, Any]:
-        assert len(self.lora_weights_lst) == 1
-        self.to_device(self.input.device)
-
-        _, num_tokens, _, _ = self.metadata()
-        # Sanity check shapes
-        i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[
-            0].shape, self.output.shape
-        # Expected input shape [num_tokens, hidden_size]
-        assert len(i_shape) == 2
-        assert i_shape[0] == num_tokens
-        hidden_size = i_shape[1]
-        # Expected lora weight shape [num_loras, lora_rank, hidden_size]
-        assert len(lw_shape) == 3
-        assert lw_shape[2] == hidden_size
-        lora_rank = lw_shape[1]
-        # Expected output shape [num_tokens, lora_rank]
-        assert len(o_shape) == 2
-        assert o_shape == (num_tokens, lora_rank)
-
-        return {
-            'inputs': self.input,
-            'lora_a_weights': self.lora_weights_lst[0],
-            'output_tensor': self.output,
-            'lora_indices_tensor': self.token_lora_mapping,
-            'scaling': 1.0
-        }
-
-    def as_bgmv_expand_kwargs(self, add_inputs: bool):
-        assert len(self.lora_weights_lst) == 1
-        self.to_device(self.input.device)
-
-        _, num_tokens, _, _ = self.metadata()
-        # Sanity check shapes
-        i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[
-            0].shape, self.output.shape
-        # Expected input shape [num_tokens, lora_rank]
-        assert len(i_shape) == 2
-        assert i_shape[0] == num_tokens
-        lora_rank = i_shape[1]
-        # Expected lora weight shape [num_loras, hidden_size, lora_rank]
-        assert len(lw_shape) == 3
-        assert lw_shape[2] == lora_rank
-        hidden_size = lw_shape[1]
-        # Expected output shape [num_tokens, hidden_size]
-        assert len(o_shape) == 2
-        assert o_shape == (num_tokens, hidden_size)
-
-        return {
-            'inputs': self.input,
-            'lora_b_weights': self.lora_weights_lst[0],
-            'output_tensor': self.output,
-            'lora_indices_tensor': self.token_lora_mapping,
-            'add_inputs': add_inputs
-        }
-
-    def as_bgmv_expand_slice_kwargs(self, add_inputs: bool) -> dict[str, Any]:
-
-        _, num_tokens, _, num_slices = self.metadata()
-        # Sanity check shapes
-        i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[
-            0].shape, self.output.shape
-        # Expected input shape [num_slices, num_tokens, lora_rank]
-        assert len(i_shape) == 3
-        assert i_shape[0] == num_slices
-        assert i_shape[1] == num_tokens
-        lora_rank = i_shape[2]
-        # Expected lora weight shape [num_loras, hidden_size, lora_rank]
-        assert len(lw_shape) == 3
-        assert lw_shape[2] == lora_rank
-        hidden_size = lw_shape[1]
-        # Expected output shape [num_tokens, hidden_size * num_slices]
-        assert len(o_shape) == 2
-        assert o_shape == (num_tokens, hidden_size * num_slices)
-
-        self.to_device(self.input.device)
-
-        kwargs_list = []
-        for i in range(num_slices):
-            kwargs_list.append({
-                'inputs': self.input[i],
-                'lora_b_weights': self.lora_weights_lst[i],
-                'output_tensor': self.output,
-                'lora_indices_tensor': self.token_lora_mapping,
-                'slice_offset': i * hidden_size,
-                'slice_size': hidden_size,
-                'add_inputs': add_inputs,
-            })
-        return {'kwargs_list': kwargs_list}
-
-    def as_v1_shrink_kwargs(self) -> dict[str, Any]:
-        assert self.v1_kernel_meta is not None
+    def as_lora_shrink_kwargs(self) -> dict[str, Any]:
         self.sanity_check()
         self.to_device(self.input.device)
 
@@ -737,17 +442,16 @@ class BenchmarkTensors:
             'inputs': self.input,
             'lora_a_weights': self.lora_weights_lst,
             'output_tensor': self.output,
-            'token_lora_mapping': self.v1_kernel_meta.token_lora_mapping,
+            'token_lora_mapping': self.lora_kernel_meta.token_lora_mapping,
             'token_indices_sorted_by_lora_ids':
-            self.v1_kernel_meta.token_indices_sorted_by_lora_ids,
-            'num_tokens_per_lora': self.v1_kernel_meta.num_tokens_per_lora,
-            'lora_token_start_loc': self.v1_kernel_meta.lora_token_start_loc,
-            'lora_ids': self.v1_kernel_meta.active_lora_ids,
+            self.lora_kernel_meta.token_indices_sorted_by_lora_ids,
+            'num_tokens_per_lora': self.lora_kernel_meta.num_tokens_per_lora,
+            'lora_token_start_loc': self.lora_kernel_meta.lora_token_start_loc,
+            'lora_ids': self.lora_kernel_meta.active_lora_ids,
             'scaling': 1.0,
         }
 
-    def as_v1_expand_kwargs(self, add_inputs: bool) -> dict[str, Any]:
-        assert self.v1_kernel_meta is not None
+    def as_lora_expand_kwargs(self, add_inputs: bool) -> dict[str, Any]:
         self.sanity_check()
         self.to_device(self.input.device)
 
@@ -773,12 +477,12 @@ class BenchmarkTensors:
             'inputs': self.input,
             'lora_b_weights': self.lora_weights_lst,
             'output_tensor': self.output,
-            'token_lora_mapping': self.v1_kernel_meta.token_lora_mapping,
+            'token_lora_mapping': self.lora_kernel_meta.token_lora_mapping,
             'token_indices_sorted_by_lora_ids':
-            self.v1_kernel_meta.token_indices_sorted_by_lora_ids,
-            'num_tokens_per_lora': self.v1_kernel_meta.num_tokens_per_lora,
-            'lora_token_start_loc': self.v1_kernel_meta.lora_token_start_loc,
-            'lora_ids': self.v1_kernel_meta.active_lora_ids,
+            self.lora_kernel_meta.token_indices_sorted_by_lora_ids,
+            'num_tokens_per_lora': self.lora_kernel_meta.num_tokens_per_lora,
+            'lora_token_start_loc': self.lora_kernel_meta.lora_token_start_loc,
+            'lora_ids': self.lora_kernel_meta.active_lora_ids,
             'offset_start': 0,
             'add_inputs': add_inputs,
         }
@@ -791,20 +495,10 @@ class BenchmarkTensors:
         else:
             assert add_inputs is not None
 
-        if op_type == OpType.SGMV_SHRINK:
-            return self.as_sgmv_shrink_kwargs()
-        if op_type == OpType.SGMV_EXPAND:
-            return self.as_sgmv_expand_kwargs(add_inputs)
-        if op_type == OpType.BGMV_SHRINK:
-            return self.as_bgmv_shrink_kwargs()
-        if op_type == OpType.BGMV_EXPAND:
-            return self.as_bgmv_expand_kwargs(add_inputs)
-        if op_type == OpType.BGMV_EXPAND_SLICE:
-            return self.as_bgmv_expand_slice_kwargs(add_inputs)
-        if op_type == OpType.V1_SHRINK:
-            return self.as_v1_shrink_kwargs()
-        if op_type == OpType.V1_EXPAND:
-            return self.as_v1_expand_kwargs(add_inputs)
+        if op_type == OpType.LORA_SHRINK:
+            return self.as_lora_shrink_kwargs()
+        if op_type == OpType.LORA_EXPAND:
+            return self.as_lora_expand_kwargs(add_inputs)
         raise ValueError(f"Unrecognized optype {self}")
 
     def test_correctness(self, op_type: OpType,
@@ -993,10 +687,6 @@ def run(args: argparse.Namespace, bench_ctxs: list[BenchmarkContext]):
     for bench_ctx in bench_ctxs:
         for seq_len in args.seq_lengths:
             bench_ops: list[OpType] = args.op_types
-            if seq_len > 1:
-                # bench only prefill ops
-                bench_ops = [op for op in args.op_types if op.is_prefill_op()]
-
             seq_len_timers = []
             for bench_op in bench_ops:
                 for num_slices in bench_op.num_slices():
@@ -1206,13 +896,13 @@ Benchmark LoRA kernels:
     {use_cuda_graph_recommendation()}
 
     list_bench example:
-        python3 benchmarks/kernels/benchmark_lora.py list_bench --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --hidden-sizes 2048 --lora-ranks 16 --num-loras 1 4 --op-types bgmv_shrink bgmv_expand sgmv_shrink sgmv_expand bgmv_expand_slice --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32
+        python3 benchmarks/kernels/benchmark_lora.py list_bench --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --hidden-sizes 2048 --lora-ranks 16 --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32
 
     model_bench example:
-        python3 benchmarks/kernels/benchmark_lora.py model_bench --models meta-llama/Llama-3-8b  --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16  --lora-ranks 16 --num-loras 1 4 --op-types bgmv_shrink bgmv_expand sgmv_shrink sgmv_expand bgmv_expand_slice --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 
+        python3 benchmarks/kernels/benchmark_lora.py model_bench --models meta-llama/Llama-3-8b  --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16  --lora-ranks 16 --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 
 
     range_bench example:
-        python3 benchmarks/kernels/benchmark_lora.py range_bench  --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16   --num-loras 1 4 --op-types bgmv_shrink bgmv_expand sgmv_shrink sgmv_expand bgmv_expand_slice --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 --hidden-sizes-start 1024 --hidden-sizes-end 4096 --hidden-sizes-increment 1024 --lora-ranks-start 8 --lora-ranks-end 24 --lora-ranks-increment 8 
+        python3 benchmarks/kernels/benchmark_lora.py range_bench  --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16   --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 --hidden-sizes-start 1024 --hidden-sizes-end 4096 --hidden-sizes-increment 1024 --lora-ranks-start 8 --lora-ranks-end 24 --lora-ranks-increment 8 
             """,  # noqa: E501
         formatter_class=argparse.RawTextHelpFormatter)
 

benchmarks/kernels/benchmark_moe.py

@@ -30,19 +30,18 @@ class BenchmarkConfig(TypedDict):
     num_stages: int
 
 
-def benchmark_config(
-    config: BenchmarkConfig,
-    num_tokens: int,
-    num_experts: int,
-    shard_intermediate_size: int,
-    hidden_size: int,
-    topk: int,
-    dtype: torch.dtype,
-    use_fp8_w8a8: bool,
-    use_int8_w8a16: bool,
-    num_iters: int = 100,
-    block_quant_shape: List[int] = None,
-) -> float:
+def benchmark_config(config: BenchmarkConfig,
+                     num_tokens: int,
+                     num_experts: int,
+                     shard_intermediate_size: int,
+                     hidden_size: int,
+                     topk: int,
+                     dtype: torch.dtype,
+                     use_fp8_w8a8: bool,
+                     use_int8_w8a16: bool,
+                     num_iters: int = 100,
+                     block_quant_shape: List[int] = None,
+                     use_deep_gemm: bool = False) -> float:
     init_dtype = torch.float16 if use_fp8_w8a8 else dtype
     x = torch.randn(num_tokens, hidden_size, dtype=dtype)
     if use_int8_w8a16:
@@ -115,22 +114,41 @@ def benchmark_config(
     def run():
         from vllm.model_executor.layers.fused_moe import override_config
         with override_config(config):
-            fused_moe(
-                x,
-                w1,
-                w2,
-                input_gating,
-                topk,
-                renormalize=True,
-                inplace=True,
-                use_fp8_w8a8=use_fp8_w8a8,
-                use_int8_w8a16=use_int8_w8a16,
-                w1_scale=w1_scale,
-                w2_scale=w2_scale,
-                a1_scale=a1_scale,
-                a2_scale=a2_scale,
-                block_shape=block_quant_shape,
-            )
+            if use_deep_gemm:
+                topk_weights, topk_ids = fused_topk(x, input_gating, topk,
+                                                    False)
+                return fused_experts(
+                    x,
+                    w1,
+                    w2,
+                    topk_weights,
+                    topk_ids,
+                    inplace=True,
+                    use_fp8_w8a8=use_fp8_w8a8,
+                    w1_scale=w1_scale,
+                    w2_scale=w2_scale,
+                    a1_scale=a1_scale,
+                    a2_scale=a2_scale,
+                    block_shape=block_quant_shape,
+                    allow_deep_gemm=True,
+                )
+            else:
+                fused_moe(
+                    x,
+                    w1,
+                    w2,
+                    input_gating,
+                    topk,
+                    renormalize=True,
+                    inplace=True,
+                    use_fp8_w8a8=use_fp8_w8a8,
+                    use_int8_w8a16=use_int8_w8a16,
+                    w1_scale=w1_scale,
+                    w2_scale=w2_scale,
+                    a1_scale=a1_scale,
+                    a2_scale=a2_scale,
+                    block_shape=block_quant_shape,
+                )
 
     # JIT compilation & warmup
     run()
@@ -366,6 +384,7 @@ class BenchmarkWorker:
         use_fp8_w8a8: bool,
         use_int8_w8a16: bool,
         block_quant_shape: List[int] = None,
+        use_deep_gemm: bool = False,
     ) -> tuple[dict[str, int], float]:
         current_platform.seed_everything(self.seed)
         dtype_str = get_config_dtype_str(dtype,
@@ -396,7 +415,8 @@ class BenchmarkWorker:
                                        use_fp8_w8a8,
                                        use_int8_w8a16,
                                        num_iters=100,
-                                       block_quant_shape=block_quant_shape)
+                                       block_quant_shape=block_quant_shape,
+                                       use_deep_gemm=use_deep_gemm)
         return config, kernel_time
 
     def tune(
@@ -411,6 +431,7 @@ class BenchmarkWorker:
         use_int8_w8a16: bool,
         search_space: list[dict[str, int]],
         block_quant_shape: list[int],
+        use_deep_gemm: bool,
     ) -> dict[str, int]:
         best_config = None
         best_time = float("inf")
@@ -436,7 +457,8 @@ class BenchmarkWorker:
                         use_fp8_w8a8,
                         use_int8_w8a16,
                         num_iters=20,
-                        block_quant_shape=block_quant_shape)
+                        block_quant_shape=block_quant_shape,
+                        use_deep_gemm=use_deep_gemm)
                 except triton.runtime.autotuner.OutOfResources:
                     # Some configurations may be invalid and fail to compile.
                     continue
@@ -550,6 +572,8 @@ def main(args: argparse.Namespace):
     else:
         batch_sizes = [args.batch_size]
 
+    use_deep_gemm = bool(args.use_deep_gemm)
+
     ray.init()
     num_gpus = int(ray.available_resources()["GPU"])
     workers = [BenchmarkWorker.remote(args.seed) for _ in range(num_gpus)]
@@ -572,10 +596,10 @@ def main(args: argparse.Namespace):
 
         start = time.time()
         configs = _distribute(
-            "tune",
-            [(batch_size, E, shard_intermediate_size, hidden_size, topk, dtype,
-              use_fp8_w8a8, use_int8_w8a16, search_space, block_quant_shape)
-             for batch_size in batch_sizes])
+            "tune", [(batch_size, E, shard_intermediate_size, hidden_size,
+                      topk, dtype, use_fp8_w8a8, use_int8_w8a16, search_space,
+                      block_quant_shape, use_deep_gemm)
+                     for batch_size in batch_sizes])
         best_configs = {
             M: sort_config(config)
             for M, config in zip(batch_sizes, configs)
@@ -589,7 +613,7 @@ def main(args: argparse.Namespace):
         outputs = _distribute(
             "benchmark",
             [(batch_size, E, shard_intermediate_size, hidden_size, topk, dtype,
-              use_fp8_w8a8, use_int8_w8a16, block_quant_shape)
+              use_fp8_w8a8, use_int8_w8a16, block_quant_shape, use_deep_gemm)
              for batch_size in batch_sizes])
 
         for batch_size, (config, kernel_time) in zip(batch_sizes, outputs):
@@ -611,6 +635,7 @@ if __name__ == "__main__":
                         type=str,
                         choices=["auto", "fp8_w8a8", "int8_w8a16"],
                         default="auto")
+    parser.add_argument("--use-deep-gemm", action="store_true")
     parser.add_argument("--seed", type=int, default=0)
     parser.add_argument("--batch-size", type=int, required=False)
     parser.add_argument("--tune", action="store_true")

benchmarks/kernels/benchmark_paged_attention.py

@@ -7,10 +7,13 @@ from typing import Optional
 import torch
 
 from vllm import _custom_ops as ops
+from vllm.logger import init_logger
 from vllm.platforms import current_platform
 from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
                         create_kv_caches_with_random)
 
+logger = init_logger(__name__)
+
 NUM_BLOCKS = 128 * 1024
 PARTITION_SIZE = 512
 PARTITION_SIZE_ROCM = 256
@@ -193,6 +196,9 @@ def main(
 
 
 if __name__ == '__main__':
+    logger.warning("This script benchmarks the paged attention kernel. "
+                   "By default this is no longer used in vLLM inference.")
+
     parser = FlexibleArgumentParser(
         description="Benchmark the paged attention kernel.")
     parser.add_argument("--version",

benchmarks/kernels/benchmark_shapes.py

@@ -75,3 +75,19 @@ WEIGHT_SHAPES = {
         [7168, 8192],
     ],
 }
+
+WEIGHT_SHAPES_MOE = {
+    "nm-testing/Mixtral-8x7B-Instruct-v0.1": [
+        [8, 2, 4096, 28672],
+        [8, 2, 14336, 4096],
+    ],
+    "nm-testing/deepseekv2-lite": [
+        [64, 6, 2048, 1408],
+    ],
+    "ibm-granite/granite-3.0-1b-a400m": [
+        [32, 8, 1024, 1024],
+    ],
+    "ibm-granite/granite-3.0-3b-a800m": [
+        [40, 8, 1024, 1536],
+    ],
+}

benchmarks/kernels/benchmark_w8a8_block_fp8.py

@@ -0,0 +1,420 @@
+# SPDX-License-Identifier: Apache-2.0
+# Adapted from sglang quantization/tuning_block_wise_kernel.py
+
+import argparse
+import json
+import multiprocessing as mp
+import os
+import time
+from datetime import datetime
+from typing import Any
+
+import torch
+import tqdm
+import triton
+
+from vllm.model_executor.layers.quantization.utils.fp8_utils import (
+    _w8a8_block_fp8_matmul)
+from vllm.platforms import current_platform
+from vllm.utils import FlexibleArgumentParser
+
+mp.set_start_method("spawn", force=True)
+
+assert current_platform.is_cuda(
+), "Only support tune w8a8 block fp8 kernel on CUDA device."
+
+DTYPE_MAP = {
+    "float32": torch.float32,
+    "float16": torch.float16,
+    "half": torch.half,
+    "bfloat16": torch.bfloat16,
+}
+
+
+def w8a8_block_matmul(
+    A: torch.Tensor,
+    B: torch.Tensor,
+    As: torch.Tensor,
+    Bs: torch.Tensor,
+    block_size: list[int],
+    config: dict[str, Any],
+    output_dtype: torch.dtype = torch.float16,
+) -> torch.Tensor:
+    """This function performs matrix multiplication with 
+    block-wise quantization.
+
+    It takes two input tensors `A` and `B` with scales `As` and `Bs`.
+    The output is returned in the specified `output_dtype`.
+
+    Args:
+        A: The input tensor, e.g., activation.
+        B: The input tensor, e.g., weight.
+        As: The per-token-group quantization scale for `A`.
+        Bs: The per-block quantization scale for `B`.
+        block_size: The block size for per-block quantization. 
+                    It should be 2-dim, e.g., [128, 128].
+        output_dytpe: The dtype of the returned tensor.
+
+    Returns:
+        torch.Tensor: The result of matmul.
+    """
+    assert len(block_size) == 2
+    block_n, block_k = block_size[0], block_size[1]
+
+    assert A.shape[-1] == B.shape[-1]
+    assert A.shape[:-1] == As.shape[:-1] and A.is_contiguous()
+    assert triton.cdiv(A.shape[-1], block_k) == As.shape[-1]
+    M = A.numel() // A.shape[-1]
+
+    assert B.ndim == 2 and B.is_contiguous() and Bs.ndim == 2
+    N, K = B.shape
+    assert triton.cdiv(N, block_n) == Bs.shape[0]
+    assert triton.cdiv(K, block_k) == Bs.shape[1]
+
+    C_shape = A.shape[:-1] + (N, )
+    C = A.new_empty(C_shape, dtype=output_dtype)
+
+    def grid(META):
+        return (triton.cdiv(M, META["BLOCK_SIZE_M"]) *
+                triton.cdiv(N, META["BLOCK_SIZE_N"]), )
+
+    if A.dtype == torch.float8_e4m3fn:
+        kernel = _w8a8_block_fp8_matmul
+    else:
+        raise RuntimeError(
+            "Currently, only support tune w8a8 block fp8 kernel.")
+
+    kernel[grid](
+        A,
+        B,
+        C,
+        As,
+        Bs,
+        M,
+        N,
+        K,
+        block_n,
+        block_k,
+        A.stride(-2),
+        A.stride(-1),
+        B.stride(1),
+        B.stride(0),
+        C.stride(-2),
+        C.stride(-1),
+        As.stride(-2),
+        As.stride(-1),
+        Bs.stride(1),
+        Bs.stride(0),
+        **config,
+    )
+
+    return C
+
+
+def get_configs_compute_bound():
+    configs = []
+    for num_stages in [2, 3, 4, 5]:
+        for block_m in [16, 32, 64, 128, 256]:
+            for block_k in [64, 128]:
+                for block_n in [32, 64, 128, 256]:
+                    for num_warps in [4, 8]:
+                        for group_size in [1, 16, 32, 64]:
+                            configs.append({
+                                "BLOCK_SIZE_M": block_m,
+                                "BLOCK_SIZE_N": block_n,
+                                "BLOCK_SIZE_K": block_k,
+                                "GROUP_SIZE_M": group_size,
+                                "num_warps": num_warps,
+                                "num_stages": num_stages,
+                            })
+    return configs
+
+
+def get_weight_shapes(tp_size):
+    # NOTE(HandH1998): The weight shapes only works for DeepSeek-V3.
+    # Modify them, if you tune for another different model.
+    # cannot TP
+    total = [
+        (512 + 64, 7168),
+        ((128 + 64) * 128, 7168),
+        (128 * (128 + 128), 512),
+        (7168, 16384),
+        (7168, 18432),
+    ]
+    # N can TP
+    n_tp = [
+        (18432 * 2, 7168),
+        ((128 + 64) * 128, 7168),
+        (128 * (128 + 128), 512),
+        (24576, 1536),
+        (12288, 7168),
+        (4096, 7168),
+    ]
+    # K can TP
+    k_tp = [(7168, 18432), (7168, 16384), (7168, 2048)]
+
+    weight_shapes = []
+    for t in total:
+        weight_shapes.append(t)
+    for n_t in n_tp:
+        new_t = (n_t[0] // tp_size, n_t[1])
+        weight_shapes.append(new_t)
+    for k_t in k_tp:
+        new_t = (k_t[0], k_t[1] // tp_size)
+        weight_shapes.append(new_t)
+    return weight_shapes
+
+
+def benchmark_config(A,
+                     B,
+                     As,
+                     Bs,
+                     block_size,
+                     config,
+                     out_dtype=torch.float16,
+                     num_iters=10):
+
+    def run():
+        w8a8_block_matmul(A, B, As, Bs, block_size, config, out_dtype)
+
+    torch.cuda.synchronize()
+    # JIT complication & warmup
+    for _ in range(5):
+        run()
+    torch.cuda.synchronize()
+
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+
+    latencies: list[float] = []
+    for i in range(num_iters):
+        torch.cuda.synchronize()
+        start_event.record()
+        run()
+        end_event.record()
+        end_event.synchronize()
+        latencies.append(start_event.elapsed_time(end_event))
+    avg = sum(latencies) / (num_iters * 10) * 1000  # us
+    return avg
+
+
+def tune(M, N, K, block_size, out_dtype, search_space, input_type):
+    factor_for_scale = 1e-2
+
+    if input_type == "fp8":
+        fp8_info = torch.finfo(torch.float8_e4m3fn)
+        fp8_max, fp8_min = fp8_info.max, fp8_info.min
+
+        A_fp32 = (
+            (torch.rand(M, K, dtype=torch.float32, device="cuda") - 0.5) * 2 *
+            fp8_max)
+        A = A_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
+
+        B_fp32 = (
+            (torch.rand(N, K, dtype=torch.float32, device="cuda") - 0.5) * 2 *
+            fp8_max)
+        B = B_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
+    else:
+        raise RuntimeError(
+            "Currently, only support tune w8a8 block fp8 kernel.")
+
+    block_n, block_k = block_size[0], block_size[1]
+    n_tiles = (N + block_n - 1) // block_n
+    k_tiles = (K + block_k - 1) // block_k
+
+    As = torch.rand(M, k_tiles, dtype=torch.float32,
+                    device="cuda") * factor_for_scale
+    Bs = (torch.rand(n_tiles, k_tiles, dtype=torch.float32, device="cuda") *
+          factor_for_scale)
+
+    best_config = None
+    best_time = float("inf")
+    for config in tqdm(search_space):
+        try:
+            kernel_time = benchmark_config(
+                A,
+                B,
+                As,
+                Bs,
+                block_size,
+                config,
+                out_dtype,
+                num_iters=10,
+            )
+        except triton.runtime.autotuner.OutOfResources:
+            # Some configurations may be invalid and fail to compile.
+            continue
+
+        if kernel_time < best_time:
+            best_time = kernel_time
+            best_config = config
+    now = datetime.now()
+    print(f"{now.ctime()}] Completed tuning for batch_size={M}")
+    assert best_config is not None
+    return best_config
+
+
+def save_configs(
+    N,
+    K,
+    block_n,
+    block_k,
+    configs,
+    save_path,
+    input_type="fp8",
+) -> None:
+    os.makedirs(save_path, exist_ok=True)
+    device_name = current_platform.get_device_name().replace(" ", "_")
+    json_file_name = (
+        f"N={N},K={K},device_name={device_name},dtype={input_type}_w8a8,"
+        f"block_shape=[{block_n},{block_k}].json")
+
+    config_file_path = os.path.join(save_path, json_file_name)
+    print(f"Writing best config to {config_file_path}...")
+
+    with open(config_file_path, "w") as f:
+        json.dump(configs, f, indent=4)
+        f.write("\n")
+
+
+def tune_on_gpu(args_dict):
+    """Run tuning on a specific GPU."""
+    gpu_id = args_dict["gpu_id"]
+    batch_sizes = args_dict["batch_sizes"]
+    weight_shapes = args_dict["weight_shapes"]
+    args = args_dict["args"]
+
+    torch.cuda.set_device(gpu_id)
+    print(f"Starting tuning on GPU {gpu_id} with batch sizes {batch_sizes}")
+
+    block_n = args.block_n
+    block_k = args.block_k
+    out_dtype = DTYPE_MAP[args.out_dtype]
+    save_path = args.save_path
+    input_type = args.input_type
+
+    search_space = get_configs_compute_bound()
+    search_space = [
+        config for config in search_space
+        if block_k % config["BLOCK_SIZE_K"] == 0
+    ]
+
+    start = time.time()
+    for shape in tqdm(weight_shapes, desc=f"GPU {gpu_id} - Shapes"):
+        N, K = shape[0], shape[1]
+        print(f"[GPU {gpu_id}] Tune for weight shape of `N: {N}, K: {K}`")
+        benchmark_results = [
+            tune(
+                batch_size,
+                N,
+                K,
+                [block_n, block_k],
+                out_dtype,
+                search_space,
+                input_type,
+            ) for batch_size in tqdm(batch_sizes,
+                                     desc=f"GPU {gpu_id} - Batch sizes")
+        ]
+        best_configs = {
+            M: config
+            for M, config in zip(batch_sizes, benchmark_results)
+        }
+        save_configs(N, K, block_n, block_k, best_configs, save_path,
+                     input_type)
+
+    end = time.time()
+    print(f"Tuning on GPU {gpu_id} took {end - start:.2f} seconds")
+
+
+def distribute_batch_sizes(batch_sizes, num_gpus):
+    """Distribute batch sizes across available GPUs."""
+    batches_per_gpu = []
+    for i in range(num_gpus):
+        start_idx = i * len(batch_sizes) // num_gpus
+        end_idx = (i + 1) * len(batch_sizes) // num_gpus
+        batches_per_gpu.append(batch_sizes[start_idx:end_idx])
+    return batches_per_gpu
+
+
+def main(args):
+    print(args)
+    num_gpus = torch.cuda.device_count()
+    if num_gpus == 0:
+        raise RuntimeError("No GPU available for tuning")
+    print(f"Found {num_gpus} GPUs for parallel tuning")
+
+    torch.cuda.init()
+
+    if args.batch_size is None:
+        batch_sizes = [
+            1,
+            2,
+            4,
+            8,
+            16,
+            24,
+            32,
+            48,
+            64,
+            96,
+            128,
+            256,
+            512,
+            1024,
+            1536,
+            2048,
+            3072,
+            4096,
+        ]
+    else:
+        batch_sizes = [args.batch_size]
+        num_gpus = 1  # If only one batch size, use only one GPU
+
+    weight_shapes = get_weight_shapes(args.tp_size)
+
+    batches_per_gpu = distribute_batch_sizes(batch_sizes, num_gpus)
+
+    process_args = []
+    for gpu_id in range(num_gpus):
+        process_args.append({
+            "gpu_id": gpu_id,
+            "batch_sizes": batches_per_gpu[gpu_id],
+            "weight_shapes":
+            weight_shapes,  # Each GPU processes all weight shapes
+            "args": args,
+        })
+
+    ctx = mp.get_context("spawn")
+    with ctx.Pool(num_gpus) as pool:
+        pool.map(tune_on_gpu, process_args)
+
+    print("Multi-GPU tuning completed")
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(
+        description="""
+Tune triton w8a8 block fp8 for DeepSeek-V3/DeepSeek-R1:
+    python3 benchmark_w8a8_block_fp8.py --tp-size 8 --input-type fp8
+Then copy to model_executor/layers/quantization/utils/configs
+        """,
+        formatter_class=argparse.RawTextHelpFormatter)
+
+    parser.add_argument("--tp-size", "-tp", type=int, default=8)
+    parser.add_argument("--input-type",
+                        type=str,
+                        choices=["fp8"],
+                        default="fp8")
+    parser.add_argument(
+        "--out-dtype",
+        type=str,
+        choices=["float32", "float16", "bfloat16", "half"],
+        default="float16",
+    )
+    parser.add_argument("--block-n", type=int, default=128)
+    parser.add_argument("--block-k", type=int, default=128)
+    parser.add_argument("--batch-size", type=int, required=False)
+    parser.add_argument("--save-path", type=str, default="./")
+    args = parser.parse_args()
+
+    main(args)

benchmarks/launch_tgi_server.sh

@@ -1,16 +0,0 @@
-#!/bin/bash
-
-PORT=8000
-MODEL=$1
-TOKENS=$2
-
-docker run -e "HF_TOKEN=$HF_TOKEN" --gpus all --shm-size 1g -p $PORT:80 \
-           -v "$PWD/data:/data" \
-           ghcr.io/huggingface/text-generation-inference:2.2.0 \
-           --model-id "$MODEL" \
-           --sharded false  \
-           --max-input-length 1024 \
-           --max-total-tokens 2048 \
-           --max-best-of 5 \
-           --max-concurrent-requests 5000 \
-           --max-batch-total-tokens "$TOKENS"

benchmarks/run_structured_output_benchmark.sh

@@ -54,6 +54,7 @@ for qps in "${QPS_VALUES[@]}"; do
   python "$SCRIPT_DIR/benchmark_serving_structured_output.py" $COMMON_PARAMS \
     --request-rate $qps \
     --result-filename "$FILENAME" \
+    --tokenizer-mode ${TOKENIZER_MODE:-"auto"} \
     --port ${PORT:-8000}
 
   echo "Completed benchmark with QPS: $qps"

cmake/cpu_extension.cmake

@@ -190,12 +190,14 @@ set(VLLM_EXT_SRC
     "csrc/cpu/cache.cpp"
     "csrc/cpu/utils.cpp"
     "csrc/cpu/layernorm.cpp"
+    "csrc/cpu/mla_decode.cpp"
     "csrc/cpu/pos_encoding.cpp"
     "csrc/cpu/torch_bindings.cpp")
 
 if (AVX512_FOUND AND NOT AVX512_DISABLED)
     set(VLLM_EXT_SRC
         "csrc/cpu/quant.cpp"
+        "csrc/cpu/shm.cpp"
         ${VLLM_EXT_SRC})
 endif()
 

cmake/external_projects/vllm_flash_attn.cmake

@@ -38,7 +38,7 @@ else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 9bfa9869829d8c593527eb34c5271d0090f7ccc9 
+          GIT_TAG dc9d410b3e2d6534a4c70724c2515f4def670a22
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn

collect_env.py

@@ -482,16 +482,28 @@ def get_pip_packages(run_lambda, patterns=None):
     if patterns is None:
         patterns = DEFAULT_PIP_PATTERNS
 
-    # People generally have `pip` as `pip` or `pip3`
-    # But here it is invoked as `python -mpip`
-    def run_with_pip(pip):
-        out = run_and_read_all(run_lambda, pip + ["list", "--format=freeze"])
+    def run_with_pip():
+        try:
+            import importlib.util
+            pip_spec = importlib.util.find_spec('pip')
+            pip_available = pip_spec is not None
+        except ImportError:
+            pip_available = False
+
+        if pip_available:
+            cmd = [sys.executable, '-mpip', 'list', '--format=freeze']
+        elif os.environ.get("UV") is not None:
+            print("uv is set")
+            cmd = ["uv", "pip", "list", "--format=freeze"]
+        else:
+            raise RuntimeError("Could not collect pip list output (pip or uv module not available)")
+
+        out = run_and_read_all(run_lambda, cmd)
         return "\n".join(line for line in out.splitlines()
                          if any(name in line for name in patterns))
 
     pip_version = 'pip3' if sys.version[0] == '3' else 'pip'
-    out = run_with_pip([sys.executable, '-mpip'])
-
+    out = run_with_pip()
     return pip_version, out
 
 

csrc/cache_kernels.cu

@@ -350,8 +350,8 @@ __global__ void concat_and_cache_mla_kernel(
 
 }  // namespace vllm
 
-// KV_T is the stored data type of kv-cache.
-// CACHE_T is the data type of key and value tensors.
+// KV_T is the data type of key and value tensors.
+// CACHE_T is the stored data type of kv-cache.
 // KV_DTYPE is the real data type of kv-cache.
 #define CALL_RESHAPE_AND_CACHE(KV_T, CACHE_T, KV_DTYPE)               \
   vllm::reshape_and_cache_kernel<KV_T, CACHE_T, KV_DTYPE>             \
@@ -393,8 +393,8 @@ void reshape_and_cache(
                              CALL_RESHAPE_AND_CACHE)
 }
 
-// KV_T is the stored data type of kv-cache.
-// CACHE_T is the data type of key and value tensors.
+// KV_T is the data type of key and value tensors.
+// CACHE_T is the stored data type of kv-cache.
 // KV_DTYPE is the real data type of kv-cache.
 #define CALL_RESHAPE_AND_CACHE_FLASH(KV_T, CACHE_T, KV_DTYPE)         \
   vllm::reshape_and_cache_flash_kernel<KV_T, CACHE_T, KV_DTYPE>       \
@@ -446,8 +446,8 @@ void reshape_and_cache_flash(
                              CALL_RESHAPE_AND_CACHE_FLASH);
 }
 
-// KV_T is the stored data type of kv-cache.
-// CACHE_T is the data type of key and value tensors.
+// KV_T is the data type of key and value tensors.
+// CACHE_T is the stored data type of kv-cache.
 // KV_DTYPE is the real data type of kv-cache.
 #define CALL_CONCAT_AND_CACHE_MLA(KV_T, CACHE_T, KV_DTYPE)              \
   vllm::concat_and_cache_mla_kernel<KV_T, CACHE_T, KV_DTYPE>            \

csrc/cpu/cache.cpp

@@ -88,6 +88,48 @@ void reshape_and_cache_cpu_impl(
 }
 };  // namespace
 
+template <typename scalar_t>
+void concat_and_cache_mla_cpu_impl(
+    const scalar_t* __restrict__ kv_c,  // [num_tokens, kv_lora_rank]
+    const scalar_t* __restrict__ k_pe,  // [num_tokens, pe_dim]
+    scalar_t* __restrict__ kv_cache,  // [num_blocks, block_size, (kv_lora_rank
+                                      // + pe_dim)]
+    const int64_t* __restrict__ slot_mapping,  // [num_tokens]
+    const int num_tokens,                      //
+    const int block_stride,                    //
+    const int entry_stride,                    //
+    const int kv_c_stride,                     //
+    const int k_pe_stride,                     //
+    const int kv_lora_rank,                    //
+    const int pe_dim,                          //
+    const int block_size                       //
+) {
+#pragma omp parallel for
+  for (int token_idx = 0; token_idx < num_tokens; ++token_idx) {
+    const int64_t slot_idx = slot_mapping[token_idx];
+    // NOTE: slot_idx can be -1 if the token is padded
+    if (slot_idx < 0) {
+      continue;
+    }
+    const int64_t block_idx = slot_idx / block_size;
+    const int64_t block_offset = slot_idx % block_size;
+
+    auto copy = [&](const scalar_t* __restrict__ src,
+                    scalar_t* __restrict__ dst, int src_stride, int dst_stride,
+                    int size, int offset) {
+      for (int i = 0; i < size; i++) {
+        const int64_t src_idx = token_idx * src_stride + i;
+        const int64_t dst_idx =
+            block_idx * block_stride + block_offset * entry_stride + i + offset;
+        dst[dst_idx] = src[src_idx];
+      }
+    };
+
+    copy(kv_c, kv_cache, kv_c_stride, block_stride, kv_lora_rank, 0);
+    copy(k_pe, kv_cache, k_pe_stride, block_stride, pe_dim, kv_lora_rank);
+  }
+}
+
 // Note: the key_caches and value_caches vectors are constant but
 // not the Tensors they contain. The vectors need to be const refs
 // in order to satisfy pytorch's C++ operator registration code.
@@ -134,6 +176,38 @@ void reshape_and_cache(torch::Tensor& key, torch::Tensor& value,
   });
 }
 
+void concat_and_cache_mla(
+    torch::Tensor& kv_c,          // [num_tokens, kv_lora_rank]
+    torch::Tensor& k_pe,          // [num_tokens, pe_dim]
+    torch::Tensor& kv_cache,      // [num_blocks, block_size, (kv_lora_rank +
+                                  // pe_dim)]
+    torch::Tensor& slot_mapping,  // [num_tokens] or [num_actual_tokens]
+    const std::string& kv_cache_dtype, torch::Tensor& scale) {
+  int num_tokens = slot_mapping.size(0);
+  int kv_lora_rank = kv_c.size(1);
+  int pe_dim = k_pe.size(1);
+  int block_size = kv_cache.size(1);
+
+  TORCH_CHECK(kv_cache.size(2) == kv_lora_rank + pe_dim);
+  TORCH_CHECK(kv_cache_dtype != "fp8");
+
+  int kv_c_stride = kv_c.stride(0);
+  int k_pe_stride = k_pe.stride(0);
+  int block_stride = kv_cache.stride(0);
+  int entry_stride = kv_cache.stride(1);
+
+  VLLM_DISPATCH_FLOATING_TYPES(
+      kv_c.scalar_type(), "concat_and_cache_mla_cpu_impl", [&] {
+        CPU_KERNEL_GUARD_IN(concat_and_cache_mla_cpu_impl)
+        concat_and_cache_mla_cpu_impl<scalar_t>(
+            kv_c.data_ptr<scalar_t>(), k_pe.data_ptr<scalar_t>(),
+            kv_cache.data_ptr<scalar_t>(), slot_mapping.data_ptr<int64_t>(),
+            num_tokens, block_stride, entry_stride, kv_c_stride, k_pe_stride,
+            kv_lora_rank, pe_dim, block_size);
+        CPU_KERNEL_GUARD_OUT(concat_and_cache_mla_cpu_impl)
+      });
+}
+
 void swap_blocks(torch::Tensor& src, torch::Tensor& dst,
                  const torch::Tensor& block_mapping) {
   TORCH_CHECK(false, "swap_blocks is unsupported on CPU.")

csrc/cpu/cpu_types_x86.hpp

@@ -78,9 +78,14 @@ struct FP16Vec16 : public Vec<FP16Vec16> {
 
   __m256i reg;
 
+  // normal load
   explicit FP16Vec16(const void* ptr)
       : reg((__m256i)_mm256_loadu_si256((__m256i*)ptr)) {}
 
+  // non-temproal load
+  explicit FP16Vec16(bool, void* ptr)
+      : reg(_mm256_stream_load_si256((__m256i*)ptr)) {}
+
   explicit FP16Vec16(const FP32Vec16&);
 
   void save(void* ptr) const { *reinterpret_cast<__m256i*>(ptr) = reg; }
@@ -110,9 +115,14 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
 
   __m256i reg;
 
+  // normal load
   explicit BF16Vec16(const void* ptr)
       : reg((__m256i)_mm256_loadu_si256((__m256i*)ptr)) {}
 
+  // non-temproal load
+  explicit BF16Vec16(bool, void* ptr)
+      : reg(_mm256_stream_load_si256((__m256i*)ptr)) {}
+
   explicit BF16Vec16(const FP32Vec16&);
 
   void save(void* ptr) const { *reinterpret_cast<__m256i*>(ptr) = reg; }
@@ -130,6 +140,8 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
 
   __m512i reg;
 
+  explicit BF16Vec32() : reg(_mm512_setzero_si512()) {}
+
   explicit BF16Vec32(const void* ptr) : reg((__m512i)_mm512_loadu_si512(ptr)) {}
 
   explicit BF16Vec32(__m512i data) : reg(data) {}
@@ -311,8 +323,13 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
 
   explicit FP32Vec16() : reg(_mm512_set1_ps(0.0)) {}
 
+  // normal load
   explicit FP32Vec16(const float* ptr) : reg(_mm512_loadu_ps(ptr)) {}
 
+  // non-temproal load
+  explicit FP32Vec16(bool, void* ptr)
+      : reg((__m512)_mm512_stream_load_si512(ptr)) {}
+
   explicit FP32Vec16(__m512 data) : reg(data) {}
 
   explicit FP32Vec16(const FP32Vec4& data)
@@ -545,6 +562,33 @@ struct INT8Vec16 : public Vec<INT8Vec16> {
     _mm_mask_storeu_epi8(ptr, mask, reg);
   }
 };
+
+struct INT8Vec64 : public Vec<INT8Vec64> {
+  constexpr static int VEC_ELEM_NUM = 64;
+  union AliasReg {
+    __m512i reg;
+    int8_t values[VEC_ELEM_NUM];
+  };
+
+  __m512i reg;
+
+  // normal load
+  explicit INT8Vec64(void* ptr) : reg(_mm512_loadu_epi8(ptr)) {}
+
+  // non-temproal load
+  explicit INT8Vec64(bool, void* ptr) : reg(_mm512_stream_load_si512(ptr)) {}
+
+  void save(void* ptr) const { _mm512_storeu_epi8(ptr, reg); }
+
+  void save(int8_t* ptr, const int elem_num) const {
+    constexpr uint64_t M = 0xFFFFFFFFFFFFFFFF;
+    __mmask64 mask = _cvtu64_mask64(M >> (64 - elem_num));
+    _mm512_mask_storeu_epi8(ptr, mask, reg);
+  }
+
+  // non-temproal save
+  void nt_save(int8_t* ptr) { _mm512_stream_si512((__m512i*)ptr, reg); }
+};
 #endif
 
 template <typename T>
@@ -655,6 +699,22 @@ inline BF16Vec16::BF16Vec16(const FP32Vec16& v) {
 
 inline void prefetch(const void* addr) { _mm_prefetch(addr, _MM_HINT_T1); }
 
+#ifdef __AVX512F__
+inline void non_temporal_save(FP16Vec16& vec, void* ptr) {
+  _mm256_stream_si256((__m256i*)ptr, vec.reg);
+}
+inline void non_temporal_save(BF16Vec32& vec, void* ptr) {
+  _mm512_stream_si512((__m512i*)ptr, vec.reg);
+}
+inline void non_temporal_save(BF16Vec16& vec, void* ptr) {
+  _mm256_stream_si256((__m256i*)ptr, vec.reg);
+}
+inline void non_temporal_save(FP32Vec16& vec, void* ptr) {
+  _mm512_stream_ps((float*)ptr, vec.reg);
+}
+#endif
+
+inline void mem_barrier() { _mm_mfence(); }
 };  // namespace vec_op
 
 #endif

csrc/cpu/mla_decode.cpp

@@ -0,0 +1,393 @@
+#include "cpu_types.hpp"
+#include <float.h>
+
+namespace {
+template <typename scalar_t>
+struct KernelVecType {
+  using qk_load_vec_type = void;
+  using qk_vec_type = void;
+  using v_load_vec_type = void;
+};
+
+template <>
+struct KernelVecType<float> {
+  using qk_load_vec_type = vec_op::FP32Vec16;
+  using qk_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::FP32Vec16;
+};
+
+template <>
+struct KernelVecType<c10::Half> {
+#if defined(__powerpc64__) || defined(__s390x__)
+  // Power and s390x architecture-specific vector types
+  using qk_load_vec_type = vec_op::FP32Vec16;
+  using qk_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::FP32Vec16;
+#else
+  // Fallback for other architectures, including x86
+  using qk_load_vec_type = vec_op::FP16Vec16;
+  using qk_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::FP16Vec16;
+#endif
+};
+
+#ifdef __AVX512BF16__
+template <>
+struct KernelVecType<c10::BFloat16> {
+  using qk_load_vec_type = vec_op::BF16Vec32;
+  using qk_vec_type = vec_op::BF16Vec32;
+  using v_load_vec_type = vec_op::BF16Vec16;
+};
+#elif defined(__aarch64__) && !defined(ARM_BF16_SUPPORT)
+// pass
+#else
+template <>
+struct KernelVecType<c10::BFloat16> {
+  using qk_load_vec_type = vec_op::BF16Vec16;
+  using qk_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::BF16Vec16;
+};
+#endif
+
+template <int HEAD_DIM, int V_HEAD_DIM, int BLOCK_SIZE, int HEAD_UNROLL,
+          typename qk_vec_type>
+void mla_decode_block_head(
+    const qk_vec_type* __restrict__ q_vecs,          // [HEAD_UNROLL, head_dim]
+    const qk_vec_type* __restrict__ k_vecs,          // [block_size, head_dim]
+    const vec_op::FP32Vec16* __restrict v_vecs_f32,  // [block_size, v_head_dim]
+    float* __restrict__ acc_out,  // [HEAD_UNROLL, v_head_dim]
+    float* __restrict__ acc_lse,  // [HEAD_UNROLL]
+    const float scale, const int num_tokens) {
+  using f32_vec_type = vec_op::FP32Vec16;
+  constexpr int QK_NUM_ELEM = qk_vec_type::VEC_ELEM_NUM;
+  constexpr int V_NUM_ELEM = f32_vec_type::VEC_ELEM_NUM;
+
+  float logits[BLOCK_SIZE][HEAD_UNROLL] = {};  // initialize to zeros
+  float max_val[HEAD_UNROLL];
+  std::fill(max_val, max_val + HEAD_UNROLL, -FLT_MAX);
+
+  f32_vec_type acc_vec[BLOCK_SIZE][HEAD_UNROLL];
+  for (int i = 0; i < HEAD_DIM; i += QK_NUM_ELEM) {
+    // load to registers
+    qk_vec_type q_vec[HEAD_UNROLL];
+
+#pragma unroll
+    for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll)
+      q_vec[unroll] =
+          qk_vec_type{q_vecs[(i + unroll * HEAD_DIM) / QK_NUM_ELEM]};
+
+    for (int block_offset = 0; block_offset < num_tokens; ++block_offset) {
+      qk_vec_type k_vec(k_vecs[(block_offset * HEAD_DIM + i) / QK_NUM_ELEM]);
+
+#pragma unroll
+      for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll)
+        vec_op::fma(acc_vec[block_offset][unroll], q_vec[unroll], k_vec);
+    }
+  }
+
+  for (int block_offset = 0; block_offset < num_tokens; ++block_offset) {
+#pragma unroll
+    for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll) {
+      const float acc = acc_vec[block_offset][unroll].reduce_sum() * scale;
+      logits[block_offset][unroll] = acc;
+      max_val[unroll] = std::max(max_val[unroll], acc);
+    }
+  }
+
+  float sum_exp[HEAD_UNROLL] = {};
+  for (int block_offset = 0; block_offset < num_tokens; ++block_offset) {
+#pragma unroll
+    for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll) {
+      const float val =
+          std::exp(logits[block_offset][unroll] - max_val[unroll]);
+      logits[block_offset][unroll] = val;
+      sum_exp[unroll] += val;
+    }
+  }
+
+  f32_vec_type this_out[V_HEAD_DIM / V_NUM_ELEM][HEAD_UNROLL];
+
+  for (int block_offset = 0; block_offset < num_tokens; ++block_offset) {
+    // load to registers
+    f32_vec_type scale_[HEAD_UNROLL];
+
+#pragma unroll
+    for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll)
+      scale_[unroll] =
+          f32_vec_type{logits[block_offset][unroll] / sum_exp[unroll]};
+
+    for (int i = 0; i < V_HEAD_DIM; i += V_NUM_ELEM) {
+      f32_vec_type v_vec(
+          v_vecs_f32[(block_offset * HEAD_DIM + i) / V_NUM_ELEM]);
+
+#pragma unroll
+      for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll)
+        vec_op::fma(this_out[i / V_NUM_ELEM][unroll], v_vec, scale_[unroll]);
+    }
+  }
+
+  // merge attention state
+  // section 2.2 in https://arxiv.org/pdf/2501.01005
+  f32_vec_type prev_scale[HEAD_UNROLL];
+  f32_vec_type curr_scale[HEAD_UNROLL];
+
+#pragma unroll
+  for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll) {
+    const float prev_lse = acc_lse[unroll];
+    const float curr_lse = std::log(sum_exp[unroll]) +
+                           max_val[unroll];  // add back max_val to get true lse
+    // softmax trick
+    const float max_lse = std::max(prev_lse, curr_lse);
+    const float prev_sum_exp = std::exp(prev_lse - max_lse);
+    const float curr_sum_exp = std::exp(curr_lse - max_lse);
+
+    const float new_sum_exp = prev_sum_exp + curr_sum_exp;
+    acc_lse[unroll] = std::log(new_sum_exp) + max_lse;
+
+    prev_scale[unroll] = f32_vec_type{prev_sum_exp / new_sum_exp};
+    curr_scale[unroll] = f32_vec_type{curr_sum_exp / new_sum_exp};
+  }
+
+  for (int i = 0; i < V_HEAD_DIM; i += V_NUM_ELEM) {
+#pragma unroll
+    for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll) {
+      f32_vec_type o_vec(acc_out + i + V_HEAD_DIM * unroll);
+      o_vec = o_vec * prev_scale[unroll] +
+              this_out[i / V_NUM_ELEM][unroll] * curr_scale[unroll];
+      o_vec.save(acc_out + i + V_HEAD_DIM * unroll);
+    }
+  }
+
+  q_vecs += HEAD_DIM / QK_NUM_ELEM * HEAD_UNROLL;
+  acc_out += V_HEAD_DIM * HEAD_UNROLL;
+}
+
+template <typename scalar_t, int HEAD_DIM, int V_HEAD_DIM, int BLOCK_SIZE,
+          typename qk_vec_type>
+void mla_decode_block(
+    const qk_vec_type* __restrict__ q_vecs,  // [num_heads, head_dim]
+    const scalar_t* __restrict__ kv_cache,   // [block_size, head_dim]
+    float* __restrict__ acc_out,             // [num_heads, v_head_dim]
+    float* __restrict__ acc_lse,             // [num_heads]
+    const int num_heads, const float scale, const int num_tokens) {
+  using qk_load_vec_type = typename KernelVecType<scalar_t>::qk_load_vec_type;
+  static_assert(
+      std::is_same<qk_vec_type,
+                   typename KernelVecType<scalar_t>::qk_vec_type>::value);
+  using v_load_vec_type = typename KernelVecType<scalar_t>::v_load_vec_type;
+  using f32_vec_type = vec_op::FP32Vec16;
+  static_assert(qk_load_vec_type::VEC_ELEM_NUM == qk_vec_type::VEC_ELEM_NUM);
+  static_assert(v_load_vec_type::VEC_ELEM_NUM == f32_vec_type::VEC_ELEM_NUM);
+  constexpr int QK_NUM_ELEM = qk_vec_type::VEC_ELEM_NUM;
+  constexpr int V_NUM_ELEM = v_load_vec_type::VEC_ELEM_NUM;
+
+  const qk_vec_type* k_vecs;
+  const f32_vec_type* v_vecs_f32;
+  float* kv_cache_f32 = nullptr;
+
+  if constexpr (!std::is_same<scalar_t, float>::value) {
+    // convert KV cache block to FP32 to reuse it across query heads and
+    // attn @ V computation, since FP16/BF16->FP32 is expensive.
+    // TODO: move malloc outside of this fn to reuse across iterations.
+    const int nbytes = BLOCK_SIZE * HEAD_DIM * sizeof(float);
+    kv_cache_f32 = static_cast<float*>(std::aligned_alloc(64, nbytes));
+
+    for (int block_offset = 0; block_offset < num_tokens; ++block_offset)
+      for (int i = 0; i < HEAD_DIM; i += V_NUM_ELEM) {
+        v_load_vec_type kv_load_vec(kv_cache + block_offset * HEAD_DIM + i);
+        f32_vec_type kv_vec_f32(kv_load_vec);
+        kv_vec_f32.save(kv_cache_f32 + block_offset * HEAD_DIM + i);
+      }
+
+    if constexpr (std::is_same<qk_load_vec_type, qk_vec_type>::value) {
+      // for AVX512_BF16, Q @ K.T uses BF16 for K (no conversion)
+      // NOTE: in this case, we only need to convert the V section to FP32.
+      // But for simplicity, we will convert the whole KV block to FP32.
+      k_vecs = reinterpret_cast<const qk_vec_type*>(kv_cache);
+    } else {
+      k_vecs = reinterpret_cast<const qk_vec_type*>(kv_cache_f32);
+    }
+
+    // attn @ V always use FP32 for V, since attn is FP32.
+    v_vecs_f32 = reinterpret_cast<const f32_vec_type*>(kv_cache_f32);
+
+  } else {
+    // KV cache is FP32. don't need to do anything.
+    k_vecs = reinterpret_cast<const qk_vec_type*>(kv_cache);
+    v_vecs_f32 = reinterpret_cast<const f32_vec_type*>(kv_cache);
+  }
+
+  // compute 2 heads at the same time to improve ILP and
+  // take advantage of register cache for K and V.
+  constexpr int HEAD_UNROLL = 2;
+  for (int iter = 0; iter < num_heads / HEAD_UNROLL; ++iter) {
+    mla_decode_block_head<HEAD_DIM, V_HEAD_DIM, BLOCK_SIZE, HEAD_UNROLL>(
+        q_vecs, k_vecs, v_vecs_f32, acc_out, acc_lse, scale, num_tokens);
+
+    q_vecs += HEAD_UNROLL * HEAD_DIM / QK_NUM_ELEM;
+    acc_out += HEAD_UNROLL * V_HEAD_DIM;
+    acc_lse += HEAD_UNROLL;
+  }
+
+  // take care of the remaining heads
+  for (int iter = 0; iter < num_heads % HEAD_UNROLL; ++iter) {
+    mla_decode_block_head<HEAD_DIM, V_HEAD_DIM, BLOCK_SIZE, 1>(
+        q_vecs, k_vecs, v_vecs_f32, acc_out, acc_lse, scale, num_tokens);
+
+    q_vecs += HEAD_DIM / QK_NUM_ELEM;
+    acc_out += V_HEAD_DIM;
+    acc_lse += 1;
+  }
+
+  if (kv_cache_f32 != nullptr) {
+    std::free(kv_cache_f32);
+  }
+}
+}  // namespace
+
+template <typename scalar_t, int HEAD_DIM, int V_HEAD_DIM, int BLOCK_SIZE>
+void mla_decode_kvcache_cpu_impl(
+    scalar_t* __restrict__ out,             // [num_seqs, num_heads, v_head_dim]
+    const scalar_t* __restrict__ q,         // [num_seqs, num_heads, head_dim]
+    const scalar_t* __restrict__ kv_cache,  // [num_blocks, block_size,
+                                            // head_dim]
+    const int num_heads, const float scale,
+    const int* __restrict__ block_tables,  // [num_seqs, max_num_blocks_per_seq]
+    const int* __restrict__ seq_lens,      // [num_seqs]
+    const int max_num_blocks_per_seq, const int o_stride, const int q_stride,
+    const int kv_stride, const int num_seqs) {
+  using qk_load_vec_type = typename KernelVecType<scalar_t>::qk_load_vec_type;
+  using qk_vec_type = typename KernelVecType<scalar_t>::qk_vec_type;
+  constexpr int QK_NUM_ELEM = qk_vec_type::VEC_ELEM_NUM;
+
+  // shared across threads
+  const int max_threads = omp_get_max_threads();
+  const int acc_out_nbytes =
+      max_threads * num_heads * V_HEAD_DIM * sizeof(float);
+  float* acc_out = static_cast<float*>(std::aligned_alloc(64, acc_out_nbytes));
+  std::vector<float> acc_lse(max_threads * num_heads);
+
+  // allocate memory to pre-convert query to FP32 later
+  float* q_f32;
+  constexpr bool PRE_CONVERT_QUERY =
+      !std::is_same<scalar_t, float>::value &&
+      std::is_same<qk_vec_type, vec_op::FP32Vec16>::value;
+  if constexpr (PRE_CONVERT_QUERY) {
+    const int q_f32_nbytes = num_heads * HEAD_DIM * sizeof(float);
+    q_f32 = static_cast<float*>(std::aligned_alloc(64, q_f32_nbytes));
+  }
+
+#pragma omp parallel
+  {
+    const int num_threads = omp_get_num_threads();
+    const int thread_id = omp_get_thread_num();
+    float* __restrict__ acc_out_thread =
+        acc_out + thread_id * num_heads * V_HEAD_DIM;
+    float* __restrict__ acc_lse_thread = acc_lse.data() + thread_id * num_heads;
+
+    for (int seq_idx = 0; seq_idx < num_seqs; ++seq_idx) {
+      // reset accumulator
+      std::fill(acc_out_thread, acc_out_thread + num_heads * V_HEAD_DIM, 0.0f);
+      std::fill(acc_lse_thread, acc_lse_thread + num_heads, -FLT_MAX);
+
+      const int seq_len = seq_lens[seq_idx];
+      const int block_num = (seq_len + BLOCK_SIZE - 1) / BLOCK_SIZE;
+      const int last_block_size = seq_len - (block_num - 1) * BLOCK_SIZE;
+
+      const qk_vec_type* q_vecs;
+      if constexpr (PRE_CONVERT_QUERY) {
+// pre-convert query to FP32 since FP16/BF16->FP32 is slow.
+#pragma omp for
+        for (int i = 0; i < num_heads * HEAD_DIM; i += QK_NUM_ELEM) {
+          qk_load_vec_type q_load_vec(q + seq_idx * q_stride + i);
+          qk_vec_type q_vec(q_load_vec);
+          q_vec.save(q_f32 + i);
+        }
+        q_vecs = reinterpret_cast<const qk_vec_type*>(q_f32);
+      } else {
+        q_vecs = reinterpret_cast<const qk_vec_type*>(q + seq_idx * q_stride);
+      }
+
+#pragma omp for
+      for (int block_idx = 0; block_idx < block_num; ++block_idx) {
+        const int physical_block_idx =
+            block_tables[seq_idx * max_num_blocks_per_seq + block_idx];
+        const int num_tokens =
+            block_idx < block_num - 1 ? BLOCK_SIZE : last_block_size;
+
+        mla_decode_block<scalar_t, HEAD_DIM, V_HEAD_DIM, BLOCK_SIZE>(
+            q_vecs, kv_cache + physical_block_idx * kv_stride, acc_out_thread,
+            acc_lse_thread, num_heads, scale, num_tokens);
+      }
+
+// merge attention states across threads
+// section 2.2 in https://arxiv.org/pdf/2501.01005
+// each thread is responsible for 1 head
+#pragma omp for
+      for (int head_idx = 0; head_idx < num_heads; ++head_idx) {
+        float* acc_lse_head = acc_lse.data() + head_idx;
+        float* acc_out_head = acc_out + head_idx * V_HEAD_DIM;
+
+        float max_val = -FLT_MAX;
+        for (int thread_id_ = 0; thread_id_ < num_threads; ++thread_id_) {
+          max_val = std::max(max_val, acc_lse_head[thread_id_ * num_heads]);
+        }
+
+        float sum_exp = 0.0f;
+        for (int thread_id_ = 0; thread_id_ < num_threads; ++thread_id_) {
+          float val = std::exp(acc_lse_head[thread_id_ * num_heads] - max_val);
+          acc_lse_head[thread_id_ * num_heads] = val;
+          sum_exp += val;
+        }
+
+        float inv_sum = 1.0f / sum_exp;
+        float out_head[V_HEAD_DIM] = {};
+        for (int thread_id_ = 0; thread_id_ < num_threads; ++thread_id_) {
+          float scale_ = acc_lse_head[thread_id_ * num_heads] * inv_sum;
+          for (int i = 0; i < V_HEAD_DIM; ++i) {
+            out_head[i] +=
+                acc_out_head[thread_id_ * num_heads * V_HEAD_DIM + i] * scale_;
+          }
+        }
+
+        for (int i = 0; i < V_HEAD_DIM; ++i) {
+          vec_op::storeFP32(out_head[i], out + seq_idx * o_stride +
+                                             head_idx * V_HEAD_DIM + i);
+        }
+      }
+    }
+  }
+  if (PRE_CONVERT_QUERY) {
+    std::free(q_f32);
+  }
+  std::free(acc_out);
+}
+
+void mla_decode_kvcache(torch::Tensor& out, torch::Tensor& query,
+                        torch::Tensor& kv_cache, double scale,
+                        torch::Tensor& block_tables, torch::Tensor& seq_lens) {
+  const int num_seqs = query.size(0);
+  const int num_heads = query.size(1);
+  const int head_dim = query.size(2);
+  const int block_size = kv_cache.size(1);
+  const int v_head_dim = out.size(2);
+
+  const int max_num_blocks_per_seq = block_tables.size(1);
+  const int o_stride = out.stride(0);
+  const int q_stride = query.stride(0);
+  const int kv_stride = kv_cache.stride(0);
+
+  VLLM_DISPATCH_FLOATING_TYPES(
+      query.scalar_type(), "mla_decode_kvcache_cpu_impl", [&] {
+        CPU_KERNEL_GUARD_IN(mla_decode_kvcache_cpu_impl)
+        if (head_dim == 576 && v_head_dim == 512 && block_size == 16)
+          mla_decode_kvcache_cpu_impl<scalar_t, 576, 512, 16>(
+              out.data_ptr<scalar_t>(), query.data_ptr<scalar_t>(),
+              kv_cache.data_ptr<scalar_t>(), num_heads, scale,
+              block_tables.data_ptr<int>(), seq_lens.data_ptr<int>(),
+              max_num_blocks_per_seq, o_stride, q_stride, kv_stride, num_seqs);
+        else
+          TORCH_CHECK(false, "Unsupported block size: ", block_size);
+        CPU_KERNEL_GUARD_OUT(mla_decode_kvcache_cpu_impl)
+      });
+}

csrc/cpu/shm.cpp

@@ -0,0 +1,781 @@
+#include "cpu/cpu_types.hpp"
+
+#include <fcntl.h>
+#include <sys/mman.h>
+#include <sys/stat.h>
+#include <unistd.h>
+
+namespace {
+#define MAX_SHM_RANK_NUM 8
+#define MAX_THREAD_NUM 12
+#define PER_THREAD_SHM_BUFFER_BYTES (4 * 1024 * 1024)
+#define MIN_THREAD_PROCESS_SIZE (8 * 1024)
+#define MAX_P2P_SEND_TENSOR_NUM 8
+
+template <typename scalar_t>
+struct KernelVecType {
+  using scalar_vec_t = void;
+};
+
+template <>
+struct KernelVecType<float> {
+  using scalar_vec_t = vec_op::FP32Vec16;
+};
+
+template <>
+struct KernelVecType<c10::BFloat16> {
+  using scalar_vec_t = vec_op::BF16Vec16;
+};
+
+template <>
+struct KernelVecType<c10::Half> {
+  using scalar_vec_t = vec_op::FP16Vec16;
+};
+
+enum class ThreadSHMStat : char { THREAD_READY = 0, SHM_DATA_READY, DONE };
+
+struct ThreadSHMContext {
+  volatile ThreadSHMStat thread_stats[MAX_SHM_RANK_NUM];
+  int thread_id;
+  int thread_num;
+  int rank;
+  int group_size;
+  size_t _spinning_count;
+  int swizzled_ranks[MAX_SHM_RANK_NUM];
+  void* thread_shm_ptrs[MAX_SHM_RANK_NUM];
+  ThreadSHMContext* shm_contexts[MAX_SHM_RANK_NUM];
+
+  ThreadSHMContext(const int thread_id, const int thread_num, const int rank,
+                   const int group_size, void* thread_shm_ptr)
+      : thread_id(thread_id),
+        thread_num(thread_num),
+        rank(rank),
+        group_size(group_size),
+        _spinning_count(0) {
+    static_assert(sizeof(ThreadSHMContext) % 64 == 0);
+    TORCH_CHECK(group_size <= MAX_SHM_RANK_NUM);
+    TORCH_CHECK((size_t)this % 64 == 0);
+    TORCH_CHECK((size_t)thread_shm_ptr % 64 == 0);
+    for (int i = 0; i < MAX_SHM_RANK_NUM; ++i) {
+      shm_contexts[i] = nullptr;
+      thread_shm_ptrs[i] = nullptr;
+      swizzled_ranks[i] = (i + rank) % group_size;
+      thread_stats[i] = ThreadSHMStat::DONE;
+    }
+    set_context(rank, this, thread_shm_ptr);
+  }
+
+  void set_context(int rank, ThreadSHMContext* ptr, void* thread_shm_ptr) {
+    TORCH_CHECK(rank < MAX_SHM_RANK_NUM);
+    TORCH_CHECK(ptr);
+    TORCH_CHECK(thread_shm_ptr);
+    TORCH_CHECK_EQ(ptr->thread_num, thread_num);
+    TORCH_CHECK_EQ(ptr->thread_id, thread_id);
+    shm_contexts[rank] = ptr;
+    thread_shm_ptrs[rank] = thread_shm_ptr;
+  }
+
+  template <typename T>
+  T* get_thread_shm_ptr(int rank) {
+    return reinterpret_cast<T*>(thread_shm_ptrs[rank]);
+  }
+
+  int get_swizzled_rank(int idx) { return swizzled_ranks[idx]; }
+
+  void wait_for_all(ThreadSHMStat prev_stat) {
+    for (int idx = 0; idx < group_size; ++idx) {
+      int rank = get_swizzled_rank(idx);
+      while (thread_stats[rank] == prev_stat) {
+        ++_spinning_count;
+        _mm_pause();
+      }
+    }
+    vec_op::mem_barrier();
+  }
+
+  void wait_for_one(int rank, ThreadSHMStat prev_stat) {
+    while (thread_stats[rank] == prev_stat) {
+      ++_spinning_count;
+      _mm_pause();
+    }
+    vec_op::mem_barrier();
+  }
+
+  void set_thread_stat(ThreadSHMStat stat) {
+    for (int idx = 0; idx < group_size; ++idx) {
+      int rank = get_swizzled_rank(idx);
+      shm_contexts[rank]->thread_stats[this->rank] = stat;
+    }
+  }
+
+  void set_thread_stat(int target_rank, ThreadSHMStat stat) {
+    for (int idx = 0; idx < group_size; ++idx) {
+      int rank = get_swizzled_rank(idx);
+      shm_contexts[rank]->thread_stats[target_rank] = stat;
+    }
+  }
+
+  // barrier for all ranks in the group, used for all2all ops
+  // DONE -> THREAD_READY -> SHM_DATA_READY -> DONE -> ...
+  void barrier(ThreadSHMStat next_stat) {
+    if (next_stat == ThreadSHMStat::THREAD_READY) {
+      set_thread_stat(ThreadSHMStat::THREAD_READY);
+      wait_for_all(ThreadSHMStat::DONE);
+    } else if (next_stat == ThreadSHMStat::SHM_DATA_READY) {
+      set_thread_stat(ThreadSHMStat::SHM_DATA_READY);
+      wait_for_all(ThreadSHMStat::THREAD_READY);
+    } else if (next_stat == ThreadSHMStat::DONE) {
+      set_thread_stat(ThreadSHMStat::DONE);
+      wait_for_all(ThreadSHMStat::SHM_DATA_READY);
+    } else {
+      TORCH_CHECK(false, "Invalid next_stat to barrier.");
+    }
+  }
+
+  std::string to_string() const {
+    std::stringstream ss;
+    ss << "SHMContext:";
+    ss << "\nrank: " << rank;
+    ss << "\ngroup_size: " << group_size;
+    ss << "\nthread_num: " << thread_num;
+    ss << "\nthread_id: " << thread_id;
+
+    ss << "\nshm_ctx_stat_loop_seq: [";
+    for (int i = 0; i < group_size; ++i) {
+      ss << swizzled_ranks[i] << ", ";
+    }
+    ss << "]";
+
+    ss << "\nshm_contexts: [";
+    for (int i = 0; i < group_size; ++i) {
+      if (shm_contexts[i]) {
+        ss << shm_contexts[i]->rank << ", ";
+      }
+    }
+    ss << "]";
+
+    return ss.str();
+  }
+};
+
+class SHMManager {
+ public:
+  explicit SHMManager(const std::string& name, const int rank,
+                      const int group_size)
+      : _rank(rank),
+        _group_size(group_size),
+        _thread_num(std::min(torch::get_num_threads(), MAX_THREAD_NUM)),
+        _shm_names({""}),
+        _shared_mem_ptrs({nullptr}),
+        _shm_ctx(nullptr) {
+    _shm_names[rank] = get_shm_name(name, rank);
+    _shared_mem_ptrs[rank] = init_shm(rank);
+    _shm_ctx = reinterpret_cast<ThreadSHMContext*>(_shared_mem_ptrs[rank]);
+
+    for (int i = 0; i < _thread_num; ++i) {
+      ThreadSHMContext* ctx = new (_shm_ctx + i)
+          ThreadSHMContext(i, _thread_num, _rank, _group_size,
+                           compute_thread_shm_ptr(_shm_ctx, i));
+    }
+  }
+
+  void join(const std::string& name) {
+    for (int rank_idx = 0; rank_idx < _group_size; ++rank_idx) {
+      if (rank_idx != _rank) {
+        TORCH_CHECK(_shm_names[rank_idx].empty());
+        TORCH_CHECK(_shared_mem_ptrs[rank_idx] == nullptr);
+        _shm_names[rank_idx] = get_shm_name(name, rank_idx);
+        _shared_mem_ptrs[rank_idx] = init_shm(rank_idx);
+        ThreadSHMContext* target_ctx =
+            reinterpret_cast<ThreadSHMContext*>(_shared_mem_ptrs[rank_idx]);
+        for (int thread_idx = 0; thread_idx < _thread_num; ++thread_idx) {
+          _shm_ctx[thread_idx].set_context(
+              rank_idx, target_ctx + thread_idx,
+              compute_thread_shm_ptr(target_ctx, thread_idx));
+        }
+      }
+    }
+  }
+
+  ~SHMManager() { destroy_shm(); }
+
+  ThreadSHMContext* get_shm_ctx() const { return _shm_ctx; }
+
+  static std::string get_shm_name(const std::string& name, int rank) {
+    return name + "_" + std::to_string(rank);
+  }
+
+  static int64_t create_singleton_instance(const std::string& name,
+                                           const int group_size,
+                                           const int rank) {
+    std::lock_guard<std::mutex> guard(SingletonInstancesLock);
+    SingletonInstances.emplace_back(
+        std::make_unique<SHMManager>(name, rank, group_size));
+    return static_cast<int64_t>(SingletonInstances.size() - 1);
+  }
+
+  static SHMManager* get_singleton_instance(int64_t handle) {
+    return SingletonInstances[handle].get();
+  }
+
+ protected:
+  static std::vector<std::unique_ptr<SHMManager>> SingletonInstances;
+  static std::mutex SingletonInstancesLock;
+
+ private:
+  static size_t round_to_alignment(size_t num) {
+    return ((num + 63) / 64) * 64;
+  }
+
+  int8_t* compute_thread_shm_ptr(ThreadSHMContext* ctx, int thread_id) {
+    int8_t* thread_shm_ptr =
+        reinterpret_cast<int8_t*>(ctx) +
+        round_to_alignment(_thread_num * sizeof(ThreadSHMContext));
+    return thread_shm_ptr +
+           thread_id * round_to_alignment(PER_THREAD_SHM_BUFFER_BYTES);
+  }
+
+  size_t compute_shm_size() {
+    const size_t rounded_rank_buffer_size =
+        round_to_alignment(PER_THREAD_SHM_BUFFER_BYTES) * _thread_num;
+    const size_t rounded_thread_shm_ctx_size =
+        round_to_alignment(_thread_num * sizeof(ThreadSHMContext));
+    const size_t shm_size =
+        rounded_thread_shm_ctx_size + rounded_rank_buffer_size;
+    return shm_size;
+  }
+
+  void* init_shm(int target_rank) {
+    const std::string& shm_name = _shm_names[target_rank];
+    const int local_rank = _rank;
+    const size_t shm_size = compute_shm_size();
+
+    int fd = -1;
+    if (local_rank == target_rank) {
+      fd = shm_open(shm_name.c_str(), O_CREAT | O_EXCL | O_RDWR,
+                    S_IRUSR | S_IWUSR);
+
+      if (fd == -1)
+        TORCH_CHECK(false, "create shm in SHMManager failed. errno: " +
+                               std::to_string(errno));
+
+      if (ftruncate(fd, shm_size) == -1)
+        TORCH_CHECK(false, "ftruncate in SHMManager failed. errno: " +
+                               std::to_string(errno));
+    } else {
+      fd = shm_open(shm_name.c_str(), O_RDWR, S_IRUSR | S_IWUSR);
+
+      if (fd == -1)
+        TORCH_CHECK(false, "open shm in SHMManager failed. errno: " +
+                               std::to_string(errno));
+    }
+
+    void* shm_ptr = mmap(nullptr, shm_size, PROT_READ | PROT_WRITE,
+                         MAP_SHARED | MAP_POPULATE, fd, 0);
+
+    if (shm_ptr == MAP_FAILED) {
+      TORCH_CHECK(false,
+                  "mmap in SHMManager failed. errno: " + std::to_string(errno));
+    }
+
+    if (close(fd) != 0) {
+      TORCH_CHECK(
+          false, "close in SHMManager failed. errno: " + std::to_string(errno));
+    }
+
+    TORCH_CHECK((size_t)shm_ptr % 64 == 0);
+
+    return shm_ptr;
+  }
+
+  void destroy_shm() {
+    std::stringstream ss;
+    ss << "local rank " << _rank << ": [";
+    for (int thread_id = 0; thread_id < _thread_num; ++thread_id) {
+      ss << _shm_ctx[thread_id]._spinning_count << ", ";
+    }
+    ss << "]\n";
+
+    for (int i = 0; i < MAX_SHM_RANK_NUM; ++i) {
+      if (_shared_mem_ptrs[i] != nullptr) {
+        munmap(_shared_mem_ptrs[i], compute_shm_size());
+      }
+
+      if (!_shm_names[i].empty()) {
+        shm_unlink(_shm_names[i].c_str());
+      }
+    }
+  }
+
+  int _rank;
+  int _group_size;
+  int _thread_num;
+  std::array<std::string, MAX_SHM_RANK_NUM> _shm_names;
+  std::array<void*, MAX_SHM_RANK_NUM> _shared_mem_ptrs;
+  ThreadSHMContext* _shm_ctx;
+};
+
+namespace shm_cc_ops {
+template <typename scalar_t, typename F>
+void shm_cc_loop(ThreadSHMContext* ctx, int64_t elem_num, F&& inner_func) {
+  int thread_num = ctx->thread_num;
+  int64_t total_bytes = elem_num * sizeof(scalar_t);
+  int64_t total_units_num =
+      (total_bytes + MIN_THREAD_PROCESS_SIZE - 1) / MIN_THREAD_PROCESS_SIZE;
+  int64_t per_thread_units_num =
+      (total_units_num + thread_num - 1) / thread_num;
+  int64_t per_unit_elem_num = MIN_THREAD_PROCESS_SIZE / sizeof(scalar_t);
+  int64_t max_per_thread_iteration_elem_num =
+      PER_THREAD_SHM_BUFFER_BYTES / sizeof(scalar_t);
+  int64_t per_thread_elem_num = per_unit_elem_num * per_thread_units_num;
+
+#pragma omp parallel for schedule(static, 1)
+  for (int i = 0; i < thread_num; ++i) {
+    int64_t offset = i * per_thread_elem_num;
+    int64_t end = std::min(elem_num, offset + per_thread_elem_num);
+    int64_t curr_elem_num =
+        std::min(max_per_thread_iteration_elem_num, end - offset);
+    ThreadSHMContext* thread_ctx = ctx + i;
+
+    while (curr_elem_num > 0) {
+      inner_func(thread_ctx, offset, curr_elem_num);
+
+      offset += max_per_thread_iteration_elem_num;
+      curr_elem_num = std::min(max_per_thread_iteration_elem_num, end - offset);
+    }
+  }
+}
+};  // namespace shm_cc_ops
+
+namespace shm_cc_ops {
+
+void memcpy_from_shm(void* dst, void* src, const int64_t bytes) {
+  const int64_t aligned_bytes = ((bytes >> 6) << 6);  // 64 bytes aligned
+  int64_t i = 0;
+#pragma GCC unroll 4
+  for (; i < aligned_bytes; i += 64) {
+    vec_op::INT8Vec64 data(
+        true, (int8_t*)src + i);  // stream loading shm to avoid caching
+    data.save((int8_t*)dst + i);
+  }
+  if (aligned_bytes < bytes) {
+    vec_op::INT8Vec64 data(true, (int8_t*)src + aligned_bytes);
+    data.save((int8_t*)dst + aligned_bytes, bytes - aligned_bytes);
+  }
+}
+
+void memcpy_to_shm(void* dst, void* src, const int64_t bytes) {
+#pragma GCC unroll 4
+  for (int64_t i = 0; i < bytes; i += 64) {
+    vec_op::INT8Vec64 data((int8_t*)src + i);
+    data.nt_save((int8_t*)dst + i);
+  }
+}
+
+void memcpy(void* dst, void* src, const int64_t bytes) {
+  const int64_t aligned_bytes = ((bytes >> 6) << 6);  // 64 bytes aligned
+  int64_t i = 0;
+#pragma GCC unroll 4
+  for (; i < aligned_bytes; i += 64) {
+    vec_op::INT8Vec64 data((int8_t*)src + i);
+    data.save((int8_t*)dst + i);
+  }
+  if (aligned_bytes < bytes) {
+    vec_op::INT8Vec64 data((int8_t*)src + aligned_bytes);
+    data.save((int8_t*)dst + aligned_bytes, bytes - aligned_bytes);
+  }
+}
+
+template <typename scalar_t, int RANKS>
+void all_reduce_sum_impl(ThreadSHMContext* ctx, scalar_t* data,
+                         size_t elem_num) {
+  CPU_KERNEL_GUARD_IN(all_reduce_sum_impl)
+  using vec_t = typename KernelVecType<scalar_t>::scalar_vec_t;
+  constexpr int64_t vec_elem_num = vec_t::get_elem_num();
+  const int worldsize = ctx->group_size;
+
+  shm_cc_ops::shm_cc_loop<scalar_t>(
+      ctx, elem_num,
+      [&](ThreadSHMContext* thread_ctx, int64_t data_offset,
+          int64_t data_elem_num) {
+        int rank = thread_ctx->rank;
+        scalar_t* thread_shm_ptr =
+            thread_ctx->get_thread_shm_ptr<scalar_t>(rank);
+        scalar_t* thread_data_ptr = data + data_offset;
+        int64_t thread_data_elem_num = data_elem_num * sizeof(scalar_t);
+
+        scalar_t* remote_data_ptrs[RANKS - 1];
+        vec_op::unroll_loop<int, RANKS - 1>([&](int idx) {
+          remote_data_ptrs[idx] = thread_ctx->get_thread_shm_ptr<scalar_t>(
+              thread_ctx->get_swizzled_rank(idx + 1));
+        });
+
+        thread_ctx->barrier(ThreadSHMStat::THREAD_READY);
+
+        shm_cc_ops::memcpy_to_shm(thread_shm_ptr, thread_data_ptr,
+                                  thread_data_elem_num);
+
+        thread_ctx->barrier(ThreadSHMStat::SHM_DATA_READY);
+
+        int64_t aligned_data_elem_num =
+            (data_elem_num / vec_elem_num) * vec_elem_num;
+        int64_t i = 0;
+#pragma GCC unroll 4
+        for (; i < aligned_data_elem_num; i += vec_elem_num) {
+          vec_t local_data(thread_data_ptr + i);  // load from cache
+          vec_op::FP32Vec16 local_data_fp32(local_data);
+          vec_op::unroll_loop<int, RANKS - 1>([&](int idx) {
+            vec_t remote_data(
+                true, remote_data_ptrs[idx] + i);  // stream load from shm
+            vec_op::FP32Vec16 remote_data_fp32(remote_data);
+            local_data_fp32 = local_data_fp32 + remote_data_fp32;  // sum reduce
+          });
+          vec_t reduced_data(local_data_fp32);
+          reduced_data.save(thread_data_ptr + i);
+        }
+
+        if (i < data_elem_num) {
+          vec_t local_data(thread_data_ptr + i);  // load from cache
+          vec_op::FP32Vec16 local_data_fp32(local_data);
+          vec_op::unroll_loop<int, RANKS - 1>([&](int idx) {
+            vec_t remote_data(
+                true, remote_data_ptrs[idx] + i);  // stream load from shm
+            vec_op::FP32Vec16 remote_data_fp32(remote_data);
+            local_data_fp32 = local_data_fp32 + remote_data_fp32;  // sum reduce
+          });
+          vec_t reduced_data(local_data_fp32);
+          reduced_data.save(thread_data_ptr + i,
+                            data_elem_num - aligned_data_elem_num);
+        }
+
+        thread_ctx->barrier(ThreadSHMStat::DONE);
+      });
+
+  return;
+}
+};  // namespace shm_cc_ops
+
+std::vector<std::unique_ptr<SHMManager>> SHMManager::SingletonInstances = {};
+std::mutex SHMManager::SingletonInstancesLock = {};
+
+template <typename scalar_t>
+void shm_allreduce_sum(ThreadSHMContext* ctx, scalar_t* data, size_t elem_num) {
+  switch (ctx->group_size) {
+    case 2:
+      shm_cc_ops::all_reduce_sum_impl<scalar_t, 2>(ctx, data, elem_num);
+      break;
+    case 3:
+      shm_cc_ops::all_reduce_sum_impl<scalar_t, 3>(ctx, data, elem_num);
+      break;
+    case 4:
+      shm_cc_ops::all_reduce_sum_impl<scalar_t, 4>(ctx, data, elem_num);
+      break;
+    case 8:
+      shm_cc_ops::all_reduce_sum_impl<scalar_t, 8>(ctx, data, elem_num);
+      break;
+    default:
+      TORCH_CHECK(false,
+                  "Invalid world size: " + std::to_string(ctx->group_size));
+  }
+}
+
+template <typename scalar_t>
+void shm_gather_impl(ThreadSHMContext* ctx, scalar_t* data, size_t elem_num,
+                     scalar_t** outputs, const int dst) {
+  CPU_KERNEL_GUARD_IN(shm_gather_impl)
+  const int worldsize = ctx->group_size;
+  TORCH_CHECK_LT(dst, worldsize);
+  shm_cc_ops::shm_cc_loop<scalar_t>(
+      ctx, elem_num,
+      [&](ThreadSHMContext* thread_ctx, int64_t data_offset,
+          int64_t data_elem_num) {
+        int rank = thread_ctx->rank;
+        scalar_t* thread_shm_ptr =
+            thread_ctx->get_thread_shm_ptr<scalar_t>(rank);
+
+        thread_ctx->barrier(ThreadSHMStat::THREAD_READY);
+
+        shm_cc_ops::memcpy_to_shm(thread_shm_ptr, data + data_offset,
+                                  data_elem_num * sizeof(scalar_t));
+
+        thread_ctx->barrier(ThreadSHMStat::SHM_DATA_READY);
+
+        if (rank == dst) {
+          shm_cc_ops::memcpy(outputs[rank] + data_offset, data + data_offset,
+                             data_elem_num * sizeof(scalar_t));
+          for (int i = 1; i < worldsize; ++i) {
+            int src_rank = thread_ctx->get_swizzled_rank(i);
+            scalar_t* src_ptr =
+                thread_ctx->get_thread_shm_ptr<scalar_t>(src_rank);  // shm
+            scalar_t* dst_ptr = outputs[src_rank] + data_offset;
+            shm_cc_ops::memcpy_from_shm(dst_ptr, src_ptr,
+                                        data_elem_num * sizeof(scalar_t));
+          }
+        }
+
+        thread_ctx->barrier(ThreadSHMStat::DONE);
+      });
+
+  return;
+}
+
+struct MemPiece {
+  void* ptr;
+  int64_t size;
+
+  template <typename T>
+  T* data_ptr() {
+    return reinterpret_cast<T*>(ptr);
+  }
+};
+
+struct TensorListMeta {
+  int64_t tensor_bytes[MAX_P2P_SEND_TENSOR_NUM];
+  torch::ScalarType tensor_types[MAX_P2P_SEND_TENSOR_NUM];
+  int64_t tensor_num;
+  int64_t total_bytes;
+
+  TensorListMeta() : tensor_num(0), total_bytes(0) {
+    static_assert(sizeof(TensorListMeta) % 64 == 0);
+    static_assert(sizeof(TensorListMeta) <
+                  MIN_THREAD_PROCESS_SIZE);  // To ensure the metadata always
+                                             // hold by the thread 0
+    for (int i = 0; i < MAX_P2P_SEND_TENSOR_NUM; ++i) {
+      tensor_bytes[i] = 0;
+      tensor_ptrs[i] = nullptr;
+      tensor_types[i] = torch::ScalarType::Undefined;
+    }
+  }
+
+  // For send and recv
+  void bind_tensor_list(std::vector<torch::Tensor>& tensor_list) {
+    TORCH_CHECK(tensor_types[0] == torch::ScalarType::Undefined,
+                "Re-bind TensorListMeta is not allowed.")
+    TORCH_CHECK_LE(tensor_list.size(), MAX_P2P_SEND_TENSOR_NUM);
+    tensor_num = tensor_list.size();
+    int64_t bytes_sum = 0;
+    for (int i = 0; i < tensor_list.size(); ++i) {
+      torch::Tensor& t = tensor_list[i];
+      TORCH_CHECK(t.is_contiguous());
+      tensor_bytes[i] = t.nbytes();
+      tensor_types[i] = t.scalar_type();
+      tensor_ptrs[i] = t.data_ptr();
+      bytes_sum += t.nbytes();
+    }
+    total_bytes = bytes_sum;
+  }
+
+  // For recv
+  std::vector<torch::Tensor> generate_tensor_list() {
+    std::vector<torch::Tensor> tensor_list;
+    tensor_list.reserve(tensor_num);
+
+    for (int i = 0; i < tensor_num; ++i) {
+      int64_t bytes = tensor_bytes[i];
+      auto type = tensor_types[i];
+      int64_t elem_bytes = torch::elementSize(type);
+
+      TORCH_CHECK_EQ(bytes % elem_bytes, 0);
+      int64_t elem_num = bytes / elem_bytes;
+      auto options = torch::TensorOptions().dtype(type).device(torch::kCPU);
+      tensor_list.emplace_back(torch::empty({elem_num}, options));
+    }
+    return tensor_list;
+  }
+
+  MemPiece get_data(int64_t offset) {
+    for (int i = 0; i < tensor_num; ++i) {
+      if (offset < tensor_bytes[i]) {
+        return {reinterpret_cast<int8_t*>(tensor_ptrs[i]) + offset,
+                tensor_bytes[i] - offset};
+      }
+      offset -= tensor_bytes[i];
+    }
+    return {nullptr, 0};
+  }
+
+ private:
+  void* tensor_ptrs[MAX_P2P_SEND_TENSOR_NUM];
+  int8_t _padding[40];
+};
+
+void shm_send_tensor_list_impl(ThreadSHMContext* ctx,
+                               const std::vector<torch::Tensor>& tensor_list) {
+  CPU_KERNEL_GUARD_IN(shm_send_tensor_list_impl)
+  std::vector<torch::Tensor> tensor_list_with_metadata;
+  tensor_list_with_metadata.reserve(1 + tensor_list.size());
+
+  auto options = torch::TensorOptions().dtype(torch::kInt8).device(torch::kCPU);
+  tensor_list_with_metadata.emplace_back(
+      torch::empty({sizeof(TensorListMeta)}, options));
+  tensor_list_with_metadata.insert(tensor_list_with_metadata.end(),
+                                   tensor_list.begin(), tensor_list.end());
+
+  torch::Tensor& metadata_tensor = tensor_list_with_metadata[0];
+  TORCH_CHECK_EQ(metadata_tensor.nbytes(), sizeof(TensorListMeta));
+
+  TensorListMeta* metadata = new (metadata_tensor.data_ptr()) TensorListMeta();
+  metadata->bind_tensor_list(tensor_list_with_metadata);
+
+  shm_cc_ops::shm_cc_loop<int8_t>(
+      ctx, metadata->total_bytes,
+      [&](ThreadSHMContext* thread_ctx, int64_t data_offset,
+          int64_t data_elem_num) {
+        int rank = thread_ctx->rank;
+        // Wait until the receiver set the stat to DONE
+        thread_ctx->wait_for_one(rank, ThreadSHMStat::SHM_DATA_READY);
+
+        int64_t curr_shm_offset = 0;
+        while (curr_shm_offset < data_elem_num) {
+          MemPiece frag = metadata->get_data(data_offset + curr_shm_offset);
+          frag.size = std::min(frag.size, data_elem_num - curr_shm_offset);
+          shm_cc_ops::memcpy(
+              thread_ctx->get_thread_shm_ptr<int8_t>(rank) + curr_shm_offset,
+              frag.ptr, frag.size);
+          curr_shm_offset += frag.size;
+        }
+
+        thread_ctx->set_thread_stat(rank, ThreadSHMStat::SHM_DATA_READY);
+      });
+}
+
+std::vector<torch::Tensor> shm_recv_tensor_list_impl(ThreadSHMContext* ctx,
+                                                     int64_t src) {
+  CPU_KERNEL_GUARD_IN(shm_recv_tensor_list_impl)
+  auto options = torch::TensorOptions().dtype(torch::kInt8).device(torch::kCPU);
+  torch::Tensor metadata_tensor =
+      torch::empty({sizeof(TensorListMeta)}, options);
+
+  // Wait until the sender set the stat of the thread 0 to SHM_DATA_READY
+  ctx->wait_for_one(src, ThreadSHMStat::DONE);
+  shm_cc_ops::memcpy(metadata_tensor.data_ptr(),
+                     ctx->get_thread_shm_ptr<void>(src),
+                     sizeof(TensorListMeta));
+  TensorListMeta* src_metadata =
+      reinterpret_cast<TensorListMeta*>(metadata_tensor.data_ptr());
+  std::vector<torch::Tensor> tensor_list_with_metadata =
+      src_metadata->generate_tensor_list();
+
+  TensorListMeta metadata;
+  metadata.bind_tensor_list(tensor_list_with_metadata);
+  TORCH_CHECK_EQ(metadata.tensor_num, src_metadata->tensor_num);
+  TORCH_CHECK_EQ(metadata.total_bytes, src_metadata->total_bytes);
+
+  shm_cc_ops::shm_cc_loop<int8_t>(
+      ctx, metadata.total_bytes,
+      [&](ThreadSHMContext* thread_ctx, int64_t data_offset,
+          int64_t data_elem_num) {
+        // Wait until the sender set the stat to SHM_DATA_READY
+        thread_ctx->wait_for_one(src, ThreadSHMStat::DONE);
+        int64_t curr_shm_offset = 0;
+        while (curr_shm_offset < data_elem_num) {
+          MemPiece frag = metadata.get_data(data_offset + curr_shm_offset);
+          frag.size = std::min(frag.size, data_elem_num - curr_shm_offset);
+          shm_cc_ops::memcpy(
+              frag.ptr,
+              thread_ctx->get_thread_shm_ptr<int8_t>(src) + curr_shm_offset,
+              frag.size);
+          curr_shm_offset += frag.size;
+        }
+
+        thread_ctx->set_thread_stat(src, ThreadSHMStat::DONE);
+      });
+
+  std::vector<torch::Tensor> tensor_list;
+  tensor_list.reserve(metadata.tensor_num - 1);
+  tensor_list.insert(tensor_list.begin(), tensor_list_with_metadata.begin() + 1,
+                     tensor_list_with_metadata.end());
+
+  return tensor_list;
+}
+}  // namespace
+
+void shm_gather(int64_t handle, torch::Tensor& data,
+                const std::optional<std::vector<torch::Tensor>>& outputs,
+                int64_t dst) {
+  TORCH_CHECK(data.is_contiguous())
+  VLLM_DISPATCH_FLOATING_TYPES(data.scalar_type(), "shm_gather_impl", [&] {
+    CPU_KERNEL_GUARD_IN(shm_gather_impl)
+
+    if (outputs.has_value()) {
+      TORCH_CHECK_LE(outputs->size(), MAX_SHM_RANK_NUM);
+      scalar_t* output_ptrs[MAX_SHM_RANK_NUM] = {nullptr};
+      for (int i = 0; i < outputs->size(); ++i) {
+        output_ptrs[i] = outputs->at(i).data_ptr<scalar_t>();
+      }
+      shm_gather_impl(SHMManager::get_singleton_instance(handle)->get_shm_ctx(),
+                      data.data_ptr<scalar_t>(), data.numel(), output_ptrs,
+                      dst);
+    } else {
+      shm_gather_impl(SHMManager::get_singleton_instance(handle)->get_shm_ctx(),
+                      data.data_ptr<scalar_t>(), data.numel(), (scalar_t**)(0),
+                      dst);
+    }
+
+    CPU_KERNEL_GUARD_OUT(shm_gather_impl)
+  });
+}
+
+void shm_all_gather(int64_t handle, const torch::Tensor& data,
+                    torch::Tensor& output) {
+  TORCH_CHECK(data.is_contiguous())
+  TORCH_CHECK(output.is_contiguous())
+
+  const int64_t input_elem_num = data.numel();
+  const int64_t output_elem_num = output.numel();
+  TORCH_CHECK_EQ(output_elem_num % input_elem_num, 0);
+  const int world_size = output_elem_num / input_elem_num;
+
+  VLLM_DISPATCH_FLOATING_TYPES(data.scalar_type(), "shm_all_gather_impl", [&] {
+    CPU_KERNEL_GUARD_IN(shm_all_gather_impl)
+    auto ctx = SHMManager::get_singleton_instance(handle)->get_shm_ctx();
+    TORCH_CHECK_EQ(ctx->group_size, world_size);
+
+    scalar_t* output_ptrs[MAX_SHM_RANK_NUM] = {nullptr};
+    for (int i = 0; i < world_size; ++i) {
+      output_ptrs[i] = output.data_ptr<scalar_t>() + i * input_elem_num;
+    }
+    shm_gather_impl(ctx, data.data_ptr<scalar_t>(), data.numel(), output_ptrs,
+                    ctx->rank);
+    CPU_KERNEL_GUARD_OUT(shm_all_gather_impl)
+  });
+}
+
+void shm_allreduce(int64_t handle, torch::Tensor& data) {
+  TORCH_CHECK(data.is_contiguous())
+  VLLM_DISPATCH_FLOATING_TYPES(data.scalar_type(), "shm_allreduce_sum", [&] {
+    CPU_KERNEL_GUARD_IN(shm_allreduce_sum)
+    shm_allreduce_sum(SHMManager::get_singleton_instance(handle)->get_shm_ctx(),
+                      data.data_ptr<scalar_t>(), data.numel());
+    CPU_KERNEL_GUARD_OUT(shm_allreduce_sum)
+  });
+}
+
+void shm_send_tensor_list(int64_t handle,
+                          const std::vector<torch::Tensor>& tensor_list,
+                          int64_t dst) {
+  CPU_KERNEL_GUARD_IN(shm_send_tensor_list)
+  shm_send_tensor_list_impl(
+      SHMManager::get_singleton_instance(handle)->get_shm_ctx(), tensor_list);
+  CPU_KERNEL_GUARD_OUT(shm_send_tensor_list)
+}
+
+std::vector<torch::Tensor> shm_recv_tensor_list(int64_t handle, int64_t src) {
+  CPU_KERNEL_GUARD_IN(shm_recv_tensor_list)
+  auto tensor_list = shm_recv_tensor_list_impl(
+      SHMManager::get_singleton_instance(handle)->get_shm_ctx(), src);
+  CPU_KERNEL_GUARD_OUT(shm_recv_tensor_list)
+  return tensor_list;
+}
+
+int64_t init_shm_manager(const std::string& name, const int64_t group_size,
+                         const int64_t rank) {
+  return SHMManager::create_singleton_instance(name, group_size, rank);
+}
+
+std::string join_shm_manager(int64_t handle, const std::string& name) {
+  auto shm_manager = SHMManager::get_singleton_instance(handle);
+  TORCH_CHECK(shm_manager);
+  shm_manager->join(name);
+  return shm_manager->get_shm_ctx()->to_string();
+}

csrc/cpu/torch_bindings.cpp

@@ -18,6 +18,30 @@ void int8_scaled_mm_azp(torch::Tensor& c, const torch::Tensor& a,
                         const std::optional<torch::Tensor>& azp,
                         const std::optional<torch::Tensor>& bias);
 
+void mla_decode_kvcache(torch::Tensor& out, torch::Tensor& query,
+                        torch::Tensor& kv_cache, double scale,
+                        torch::Tensor& block_tables, torch::Tensor& seq_lens);
+
+int64_t init_shm_manager(const std::string& name, const int64_t group_size,
+                         const int64_t rank);
+
+std::string join_shm_manager(int64_t handle, const std::string& name);
+
+void shm_allreduce(int64_t handle, torch::Tensor& data);
+
+void shm_gather(int64_t handle, torch::Tensor& data,
+                const std::optional<std::vector<torch::Tensor>>& outputs,
+                int64_t dst);
+
+void shm_all_gather(int64_t handle, const torch::Tensor& data,
+                    torch::Tensor& output);
+
+void shm_send_tensor_list(int64_t handle,
+                          const std::vector<torch::Tensor>& tensor_list,
+                          int64_t dst);
+
+std::vector<torch::Tensor> shm_recv_tensor_list(int64_t handle, int64_t src);
+
 TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   // vLLM custom ops
 
@@ -127,6 +151,29 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "                  Tensor? azp, Tensor? bias) -> ()");
   ops.impl("cutlass_scaled_mm_azp", torch::kCPU, &int8_scaled_mm_azp);
 #endif
+
+// SHM CCL
+#ifdef __AVX512F__
+  ops.def("init_shm_manager(str name, int group_size, int rank) -> int",
+          &init_shm_manager);
+  ops.def("join_shm_manager(int handle, str name) -> str", &join_shm_manager);
+  ops.def("shm_allreduce(int handle, Tensor! data) -> ()");
+  ops.impl("shm_allreduce", torch::kCPU, &shm_allreduce);
+  ops.def(
+      "shm_gather(int handle, Tensor data, Tensor[](a!)? outputs, int dst) -> "
+      "()");
+  ops.impl("shm_gather", torch::kCPU, &shm_gather);
+  ops.def(
+      "shm_all_gather(int handle, Tensor data, Tensor! output) -> "
+      "()");
+  ops.impl("shm_all_gather", torch::kCPU, &shm_all_gather);
+  ops.def(
+      "shm_send_tensor_list(int handle, Tensor[](a) tensor_list, int dst) -> "
+      "()");
+  ops.impl("shm_send_tensor_list", torch::kCPU, &shm_send_tensor_list);
+  ops.def("shm_recv_tensor_list(int handle, int src) -> Tensor[](a)",
+          &shm_recv_tensor_list);
+#endif
 }
 
 TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cache_ops), cache_ops) {
@@ -150,6 +197,14 @@ TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cache_ops), cache_ops) {
       "                  str kv_cache_dtype,"
       "                  Tensor k_scale, Tensor v_scale) -> ()");
   cache_ops.impl("reshape_and_cache", torch::kCPU, &reshape_and_cache);
+
+  cache_ops.def(
+      "concat_and_cache_mla(Tensor kv_c, Tensor k_pe,"
+      "                     Tensor! kv_cache,"
+      "                     Tensor slot_mapping,"
+      "                     str kv_cache_dtype,"
+      "                     Tensor scale) -> ()");
+  cache_ops.impl("concat_and_cache_mla", torch::kCPU, &concat_and_cache_mla);
 }
 
 TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _utils), utils) {
@@ -157,4 +212,12 @@ TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _utils), utils) {
   utils.def("init_cpu_threads_env(str cpu_ids) -> str", &init_cpu_threads_env);
 }
 
+TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cpu), cpu_ops) {
+  cpu_ops.def(
+      "mla_decode_kvcache("
+      "   Tensor! out, Tensor query, Tensor kv_cache,"
+      "   float scale, Tensor block_tables, Tensor seq_lens) -> ()");
+  cpu_ops.impl("mla_decode_kvcache", torch::kCPU, &mla_decode_kvcache);
+}
+
 REGISTER_EXTENSION(TORCH_EXTENSION_NAME)

csrc/cpu/utils.cpp

@@ -18,7 +18,7 @@ std::string init_cpu_threads_env(const std::string& cpu_ids) {
 
 #ifndef VLLM_NUMA_DISABLED
 std::string init_cpu_threads_env(const std::string& cpu_ids) {
-  bitmask* omp_cpu_mask = numa_parse_cpustring(cpu_ids.c_str());
+  bitmask* omp_cpu_mask = numa_parse_cpustring_all(cpu_ids.c_str());
   TORCH_CHECK(omp_cpu_mask->size > 0);
   std::vector<int> omp_cpu_ids;
   omp_cpu_ids.reserve(omp_cpu_mask->size);

csrc/cuda_view.cu

@@ -0,0 +1,39 @@
+#include <torch/all.h>
+#include <torch/cuda.h>
+#include <cuda_runtime.h>
+
+// This function assumes that `cpu_tensor` is a CPU tensor allocated with pinned
+// memory, and that UVA (Unified Virtual Addressing) is enabled.
+torch::Tensor get_cuda_view_from_cpu_tensor(torch::Tensor& cpu_tensor) {
+  TORCH_CHECK(cpu_tensor.device().is_cpu(), "Input tensor must be on CPU");
+
+  // Get raw host pointer from CPU tensor
+  void* host_ptr = cpu_tensor.data_ptr();
+
+  // Get a device pointer corresponding to the pinned host memory
+  void* device_ptr = nullptr;
+  cudaError_t err = cudaHostGetDevicePointer(&device_ptr, host_ptr, 0);
+  TORCH_CHECK(err == cudaSuccess,
+              "cudaHostGetDevicePointer failed: ", cudaGetErrorString(err));
+
+  // We'll use the same sizes, strides, and dtype as the CPU tensor.
+  // TODO: check if layout is respected.
+  auto sizes = cpu_tensor.sizes();
+  auto strides = cpu_tensor.strides();
+  auto options = cpu_tensor.options().device(torch::kCUDA);
+
+  // from_blob signature: from_blob(void *data, IntArrayRef sizes, ..., Deleter,
+  // const TensorOptions &) Provide a no-op deleter. The CPU tensor holds the
+  // memory, so we don't free it here.
+  auto deleter = [](void*) {
+    // no-op, since the memory is owned by the original CPU tensor
+  };
+
+  torch::Tensor cuda_tensor =
+      torch::from_blob(device_ptr, sizes, strides, deleter, options);
+
+  TORCH_CHECK(cuda_tensor.device().is_cuda(),
+              "Resulting tensor is not on CUDA device");
+
+  return cuda_tensor;
+}

csrc/custom_all_reduce.cu

@@ -12,7 +12,7 @@ static_assert(sizeof(void*) == sizeof(fptr_t));
 
 fptr_t init_custom_ar(const std::vector<fptr_t>& fake_ipc_ptrs,
                       torch::Tensor& rank_data, int64_t rank,
-                      bool full_nvlink) {
+                      bool fully_connected) {
   int world_size = fake_ipc_ptrs.size();
   if (world_size > 8)
     throw std::invalid_argument("world size > 8 is not supported");
@@ -27,7 +27,7 @@ fptr_t init_custom_ar(const std::vector<fptr_t>& fake_ipc_ptrs,
   }
   return (fptr_t) new vllm::CustomAllreduce(ipc_ptrs, rank_data.data_ptr(),
                                             rank_data.numel(), rank, world_size,
-                                            full_nvlink);
+                                            fully_connected);
 }
 
 /**
@@ -142,3 +142,48 @@ void register_graph_buffers(fptr_t _fa,
   bytes.reserve(handles.size());
   fa->register_graph_buffers(bytes, offsets);
 }
+
+std::tuple<fptr_t, torch::Tensor> allocate_shared_buffer_and_handle(
+    int64_t size) {
+  auto device_index = c10::cuda::current_device();
+  at::DeviceGuard device_guard(at::Device(at::DeviceType::CUDA, device_index));
+  void* buffer;
+  cudaStreamCaptureMode mode = cudaStreamCaptureModeRelaxed;
+  auto stream = c10::cuda::getCurrentCUDAStream().stream();
+  AT_CUDA_CHECK(cudaThreadExchangeStreamCaptureMode(&mode));
+
+  // Allocate buffer
+#if defined(USE_ROCM)
+  // data buffers need to be "uncached" for signal on MI200
+  AT_CUDA_CHECK(
+      hipExtMallocWithFlags((void**)&buffer, size, hipDeviceMallocUncached));
+#else
+  AT_CUDA_CHECK(cudaMalloc((void**)&buffer, size));
+#endif
+  AT_CUDA_CHECK(cudaMemsetAsync(buffer, 0, size, stream));
+  AT_CUDA_CHECK(cudaStreamSynchronize(stream));
+  AT_CUDA_CHECK(cudaThreadExchangeStreamCaptureMode(&mode));
+
+  // Create IPC memhandle for the allocated buffer.
+  // Will use it in open_mem_handle.
+  auto options =
+      torch::TensorOptions().dtype(torch::kUInt8).device(torch::kCPU);
+  auto handle =
+      torch::empty({static_cast<int64_t>(sizeof(cudaIpcMemHandle_t))}, options);
+  AT_CUDA_CHECK(
+      cudaIpcGetMemHandle((cudaIpcMemHandle_t*)handle.data_ptr(), buffer));
+
+  return std::make_tuple(reinterpret_cast<fptr_t>(buffer), handle);
+}
+
+fptr_t open_mem_handle(torch::Tensor& mem_handle) {
+  void* ipc_ptr;
+  AT_CUDA_CHECK(cudaIpcOpenMemHandle(
+      (void**)&ipc_ptr, *((const cudaIpcMemHandle_t*)mem_handle.data_ptr()),
+      cudaIpcMemLazyEnablePeerAccess));
+  return reinterpret_cast<fptr_t>(ipc_ptr);
+}
+
+void free_shared_buffer(fptr_t buffer) {
+  AT_CUDA_CHECK(cudaFree(reinterpret_cast<void*>(buffer)));
+}

csrc/custom_all_reduce.cuh

@@ -5,6 +5,10 @@
 #include <cuda_fp16.h>
 #include <cuda_runtime.h>
 
+#if defined(USE_ROCM)
+typedef __hip_bfloat16 nv_bfloat16;
+#endif
+
 #include <iostream>
 #include <array>
 #include <limits>
@@ -12,6 +16,7 @@
 #include <unordered_map>
 #include <vector>
 
+namespace vllm {
 #define CUDACHECK(cmd)                                              \
   do {                                                              \
     cudaError_t e = cmd;                                            \
@@ -22,24 +27,37 @@
     }                                                               \
   } while (0)
 
-namespace vllm {
-
+// Maximal number of blocks in allreduce kernel.
 constexpr int kMaxBlocks = 36;
+
+// Default number of blocks in allreduce kernel.
+#ifndef USE_ROCM
+const int defaultBlockLimit = 36;
+CUpointer_attribute rangeStartAddrAttr = CU_POINTER_ATTRIBUTE_RANGE_START_ADDR;
+#else
+const int defaultBlockLimit = 16;
+hipPointer_attribute rangeStartAddrAttr =
+    HIP_POINTER_ATTRIBUTE_RANGE_START_ADDR;
+#endif
+
 // Counter may overflow, but it's fine since unsigned int overflow is
 // well-defined behavior.
 using FlagType = uint32_t;
+
+// Two sets of peer counters are needed for two syncs: starting and ending an
+// operation. The reason is that it's possible for peer GPU block to arrive at
+// the second sync point while the current GPU block haven't passed the first
+// sync point. Thus, peer GPU may write counter+1 while current GPU is busy
+// waiting for counter. We use alternating counter array to avoid this
+// possibility.
 struct Signal {
-  alignas(128) FlagType self_counter[kMaxBlocks][8];
-  // Two sets of peer counters are needed for two syncs. The reason is that
-  // it's possible for peer GPU block to arrive at the second sync point while
-  // the current GPU block haven't passed the first sync point. Thus, peer GPU
-  // may write counter+1 while current GPU is busy waiting for counter. We use
-  // alternating counter array to avoid this possibility.
-  alignas(128) FlagType peer_counter[2][kMaxBlocks][8];
+  alignas(128) FlagType start[kMaxBlocks][8];
+  alignas(128) FlagType end[kMaxBlocks][8];
+  alignas(128) FlagType _flag[kMaxBlocks];  // incremental flags for each rank
 };
 
 struct __align__(16) RankData {
-  const void* __restrict__ ptrs[8];
+  const void* ptrs[8];
 };
 
 struct __align__(16) RankSignals {
@@ -134,27 +152,29 @@ DINLINE O downcast(array_t<float, O::size> val) {
   }
 }
 
+#if !defined(USE_ROCM)
+
 static DINLINE void st_flag_release(FlagType* flag_addr, FlagType flag) {
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
+  #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
   asm volatile("st.release.sys.global.u32 [%1], %0;" ::"r"(flag),
                "l"(flag_addr));
-#else
+  #else
   asm volatile("membar.sys; st.volatile.global.u32 [%1], %0;" ::"r"(flag),
                "l"(flag_addr));
-#endif
+  #endif
 }
 
 static DINLINE FlagType ld_flag_acquire(FlagType* flag_addr) {
   FlagType flag;
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
+  #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
   asm volatile("ld.acquire.sys.global.u32 %0, [%1];"
                : "=r"(flag)
                : "l"(flag_addr));
-#else
+  #else
   asm volatile("ld.volatile.global.u32 %0, [%1]; membar.gl;"
                : "=r"(flag)
                : "l"(flag_addr));
-#endif
+  #endif
   return flag;
 }
 
@@ -170,37 +190,99 @@ static DINLINE FlagType ld_flag_volatile(FlagType* flag_addr) {
   return flag;
 }
 
-// is_start: whether this is the very first synchronization barrier.
-// need_fence: whether a memory fence is needed. If true, a release-acquire
-// semantic is used to enforce memory access order before and after this
-// barrier.
-template <int ngpus, bool is_start, bool need_fence = false>
-DINLINE void multi_gpu_barrier(const RankSignals& sg, Signal* self_sg,
-                               int rank) {
-  if constexpr (!is_start) __syncthreads();
-  static_assert(
-      !(is_start && need_fence));  // Start barrier shouldn't need fence.
+// This function is meant to be used as the first synchronization in the all
+// reduce kernel. Thus, it doesn't need to make any visibility guarantees for
+// prior memory accesses. Note: volatile writes will not be reordered against
+// other volatile writes.
+template <int ngpus>
+DINLINE void barrier_at_start(const RankSignals& sg, Signal* self_sg,
+                              int rank) {
+  uint32_t flag = self_sg->_flag[blockIdx.x] + 1;
   if (threadIdx.x < ngpus) {
-    // Increment the counter. Technically we only need one counter, but we use
-    // multiple per block to eliminate the need to share the counter via smem.
-    auto val = self_sg->self_counter[blockIdx.x][threadIdx.x] += 1;
+    auto peer_counter_ptr = &sg.signals[threadIdx.x]->start[blockIdx.x][rank];
+    auto self_counter_ptr = &self_sg->start[blockIdx.x][threadIdx.x];
+    // Write the expected counter value to peer and wait for correct value
+    // from peer.
+    st_flag_volatile(peer_counter_ptr, flag);
+    while (ld_flag_volatile(self_counter_ptr) != flag);
+  }
+  __syncthreads();
+  // use one thread to update flag
+  if (threadIdx.x == 0) self_sg->_flag[blockIdx.x] = flag;
+}
+
+// This function is meant to be used as the second or the final
+// synchronization barrier in the all reduce kernel. If it's the final
+// synchronization barrier, we don't need to make any visibility guarantees
+// for prior memory accesses.
+template <int ngpus, bool final_sync = false>
+DINLINE void barrier_at_end(const RankSignals& sg, Signal* self_sg, int rank) {
+  __syncthreads();
+  uint32_t flag = self_sg->_flag[blockIdx.x] + 1;
+  if (threadIdx.x < ngpus) {
+    auto peer_counter_ptr = &sg.signals[threadIdx.x]->end[blockIdx.x][rank];
+    auto self_counter_ptr = &self_sg->end[blockIdx.x][threadIdx.x];
     // Write the expected counter value to peer and wait for correct value from
     // peer.
-    auto peer_counter_ptr =
-        &sg.signals[threadIdx.x]->peer_counter[val % 2][blockIdx.x][rank];
-    auto self_counter_ptr =
-        &self_sg->peer_counter[val % 2][blockIdx.x][threadIdx.x];
-    if constexpr (need_fence) {
-      st_flag_release(peer_counter_ptr, val);
-      while (ld_flag_acquire(self_counter_ptr) != val);
+    if constexpr (!final_sync) {
+      st_flag_release(peer_counter_ptr, flag);
+      while (ld_flag_acquire(self_counter_ptr) != flag);
     } else {
-      st_flag_volatile(peer_counter_ptr, val);
-      while (ld_flag_volatile(self_counter_ptr) != val);
+      st_flag_volatile(peer_counter_ptr, flag);
+      while (ld_flag_volatile(self_counter_ptr) != flag);
     }
   }
-  if constexpr (is_start || need_fence) __syncthreads();
+  if constexpr (!final_sync) __syncthreads();
+
+  // use one thread to update flag
+  if (threadIdx.x == 0) self_sg->_flag[blockIdx.x] = flag;
 }
 
+#else
+
+template <int ngpus>
+DINLINE void barrier_at_start(const RankSignals& sg, Signal* self_sg,
+                              int rank) {
+  uint32_t flag = self_sg->_flag[blockIdx.x] + 1;
+  if (threadIdx.x < ngpus) {
+    // simultaneously write to the corresponding flag of all ranks.
+    // Latency = 1 p2p write
+    __scoped_atomic_store_n(&sg.signals[threadIdx.x]->start[blockIdx.x][rank],
+                            flag, __ATOMIC_RELAXED, __MEMORY_SCOPE_SYSTEM);
+    // wait until we got true from all ranks
+    while (__scoped_atomic_load_n(&self_sg->start[blockIdx.x][threadIdx.x],
+                                  __ATOMIC_RELAXED,
+                                  __MEMORY_SCOPE_DEVICE) < flag);
+  }
+  __syncthreads();
+  // use one thread to update flag
+  if (threadIdx.x == 0) self_sg->_flag[blockIdx.x] = flag;
+}
+
+template <int ngpus, bool final_sync = false>
+DINLINE void barrier_at_end(const RankSignals& sg, Signal* self_sg, int rank) {
+  __syncthreads();
+  uint32_t flag = self_sg->_flag[blockIdx.x] + 1;
+  if (threadIdx.x < ngpus) {
+    // simultaneously write to the corresponding flag of all ranks.
+    // Latency = 1 p2p write
+    __scoped_atomic_store_n(&sg.signals[threadIdx.x]->end[blockIdx.x][rank],
+                            flag,
+                            final_sync ? __ATOMIC_RELAXED : __ATOMIC_RELEASE,
+                            __MEMORY_SCOPE_SYSTEM);
+    // wait until we got true from all ranks
+    while (
+        __scoped_atomic_load_n(&self_sg->end[blockIdx.x][threadIdx.x],
+                               final_sync ? __ATOMIC_RELAXED : __ATOMIC_ACQUIRE,
+                               __MEMORY_SCOPE_DEVICE) < flag);
+  }
+  if constexpr (!final_sync) __syncthreads();
+  // use one thread to update flag
+  if (threadIdx.x == 0) self_sg->_flag[blockIdx.x] = flag;
+}
+
+#endif
+
 template <typename P, int ngpus, typename A>
 DINLINE P packed_reduce(const P* ptrs[], int idx) {
   A tmp = upcast(ptrs[0][idx]);
@@ -220,13 +302,13 @@ __global__ void __launch_bounds__(512, 1)
   // note: we don't reorder the address so the accumulation order is the same
   // for all ranks, ensuring bitwise identical results
   auto dp = *_dp;
-  multi_gpu_barrier<ngpus, true>(sg, self_sg, rank);
+  barrier_at_start<ngpus>(sg, self_sg, rank);
   // do the actual reduction
   for (int idx = blockIdx.x * blockDim.x + threadIdx.x; idx < size;
        idx += gridDim.x * blockDim.x) {
     ((P*)result)[idx] = packed_reduce<P, ngpus, A>((const P**)&dp.ptrs[0], idx);
   }
-  multi_gpu_barrier<ngpus, false>(sg, self_sg, rank);
+  barrier_at_end<ngpus, true>(sg, self_sg, rank);
 }
 
 template <typename P>
@@ -255,18 +337,20 @@ __global__ void __launch_bounds__(512, 1)
     tmps[i] = get_tmp_buf<P>(sg.signals[target]);
   }
   auto tmp_out = tmps[0];
-  multi_gpu_barrier<ngpus, true>(sg, self_sg, rank);
+  barrier_at_start<ngpus>(sg, self_sg, rank);
+
   // stage 1: reduce scatter
   for (int idx = start + tid; idx < end; idx += stride) {
     tmp_out[idx - start] = packed_reduce<P, ngpus, A>(ptrs, idx);
   }
-  multi_gpu_barrier<ngpus, false, true>(sg, self_sg, rank);
+  barrier_at_end<ngpus>(sg, self_sg, rank);
 
   // stage 2: allgather. Note: it's important to match the tid between
   // the two stages, because visibility across devices is only guaranteed
   // between threads that have the same tid. If thread i computes the sum of
-  // start + i in the first stage, then thread i also gathers start + i from all
-  // ranks.
+  // start + i in the first stage, then thread i also gathers start + i from
+  // all ranks.
+
   for (int idx = tid; idx < largest_part; idx += stride) {
 #pragma unroll
     for (int i = 0; i < ngpus; i++) {
@@ -287,21 +371,22 @@ class CustomAllreduce {
  public:
   int rank_;
   int world_size_;
-  bool full_nvlink_;
+  // Full NVLink or xGMI connection between GPUs.
+  bool fully_connected_;
 
   RankSignals sg_;
-  // Stores an map from a pointer to its peer pointters from all ranks.
+  // Stores an map from a pointer to its peer pointers from all ranks.
   std::unordered_map<void*, RankData*> buffers_;
   Signal* self_sg_;
 
   // Stores rank data from all ranks. This is mainly for cuda graph purposes.
   // For cuda graph to work, all kernel arguments must be fixed during graph
-  // capture time. However, the peer pointers are not known during graph capture
-  // time. Therefore, during capture, we increment the rank data pointer and use
-  // that as the argument to the kernel. The kernel arguments are stored in
-  // graph_unreg_buffers_. The actual peer pointers will be filled in at the
-  // memory pointed to by the pointers in graph_unreg_buffers_ when
-  // the IPC handles are exchanged between ranks.
+  // capture time. However, the peer pointers are not known during graph
+  // capture time. Therefore, during capture, we increment the rank data
+  // pointer and use that as the argument to the kernel. The kernel arguments
+  // are stored in graph_unreg_buffers_. The actual peer pointers will be
+  // filled in at the memory pointed to by the pointers in
+  // graph_unreg_buffers_ when the IPC handles are exchanged between ranks.
   //
   // The overall process looks like this:
   // 1. Graph capture.
@@ -319,17 +404,18 @@ class CustomAllreduce {
    * Signals are an array of ipc-enabled buffers from all ranks.
    * For each of the buffer, the layout is as follows:
    * | -- sizeof(Signal) -- | ------ a few MB ----- |
-   * The first section is for allreduce synchronization, and the second section
-   * is for storing the intermediate results required by some allreduce algos.
+   * The first section is for allreduce synchronization, and the second
+   * section is for storing the intermediate results required by some
+   * allreduce algos.
    *
    * Note: this class does not own any device memory. Any required buffers
    * are passed in from the constructor.
    */
   CustomAllreduce(Signal** signals, void* rank_data, size_t rank_data_sz,
-                  int rank, int world_size, bool full_nvlink = true)
+                  int rank, int world_size, bool fully_connected = true)
       : rank_(rank),
         world_size_(world_size),
-        full_nvlink_(full_nvlink),
+        fully_connected_(fully_connected),
         self_sg_(signals[rank]),
         d_rank_data_base_(reinterpret_cast<RankData*>(rank_data)),
         d_rank_data_end_(d_rank_data_base_ + rank_data_sz / sizeof(RankData)) {
@@ -361,8 +447,7 @@ class CustomAllreduce {
       void* base_ptr;
       // note: must share the base address of each allocation, or we get wrong
       // address
-      if (cuPointerGetAttribute(&base_ptr,
-                                CU_POINTER_ATTRIBUTE_RANGE_START_ADDR,
+      if (cuPointerGetAttribute(&base_ptr, rangeStartAddrAttr,
                                 (CUdeviceptr)ptr) != CUDA_SUCCESS)
         throw std::runtime_error("failed to get pointer attr");
       CUDACHECK(cudaIpcGetMemHandle(
@@ -396,11 +481,11 @@ class CustomAllreduce {
 
   // Note: when registering graph buffers, we intentionally choose to not
   // deduplicate the addresses. That means if the allocator reuses some
-  // addresses, they will be registered again. This is to account for the remote
-  // possibility of different allocation patterns between ranks. For example,
-  // rank 1 may get the same input address for the second allreduce, but rank 2
-  // got a different address. IPC handles have internal reference counting
-  // mechanism so overhead should be small.
+  // addresses, they will be registered again. This is to account for the
+  // remote possibility of different allocation patterns between ranks. For
+  // example, rank 1 may get the same input address for the second allreduce,
+  // but rank 2 got a different address. IPC handles have internal reference
+  // counting mechanism so overhead should be small.
   void register_graph_buffers(
       const std::vector<std::string>& handles,
       const std::vector<std::vector<int64_t>>& offsets) {
@@ -431,15 +516,15 @@ class CustomAllreduce {
   /**
    * Performs allreduce, assuming input has already been registered.
    *
-   * Block and grid default configs are results after careful grid search. Using
-   * 36 blocks give the best or close to the best runtime on the devices I
-   * tried: A100, A10, A30, T4, V100. You'll notice that NCCL kernels also only
-   * take a small amount of SMs. Not quite sure the underlying reason, but my
-   * guess is that too many SMs will cause contention on NVLink bus.
+   * Block and grid default configs are results after careful grid search.
+   * Using 36 blocks give the best or close to the best runtime on the devices
+   * I tried: A100, A10, A30, T4, V100. You'll notice that NCCL kernels also
+   * only take a small amount of SMs. Not quite sure the underlying reason,
+   * but my guess is that too many SMs will cause contention on NVLink bus.
    */
   template <typename T>
   void allreduce(cudaStream_t stream, T* input, T* output, int size,
-                 int threads = 512, int block_limit = 36) {
+                 int threads = 512, int block_limit = defaultBlockLimit) {
     auto d = packed_t<T>::P::size;
     if (size % d != 0)
       throw std::runtime_error(
@@ -473,13 +558,11 @@ class CustomAllreduce {
 #define KL(ngpus, name)                                                       \
   name<T, ngpus><<<blocks, threads, 0, stream>>>(ptrs, sg_, self_sg_, output, \
                                                  rank_, size);
-    // TODO(hanzhi713): Threshold is different for A100 and H100.
-    // Add per device threshold.
 #define REDUCE_CASE(ngpus)                            \
   case ngpus: {                                       \
     if (world_size_ == 2) {                           \
       KL(ngpus, cross_device_reduce_1stage);          \
-    } else if (full_nvlink_) {                        \
+    } else if (fully_connected_) {                    \
       if ((world_size_ <= 4 && bytes < 512 * 1024) || \
           (world_size_ <= 8 && bytes < 256 * 1024)) { \
         KL(ngpus, cross_device_reduce_1stage);        \
@@ -497,7 +580,8 @@ class CustomAllreduce {
       REDUCE_CASE(8)
       default:
         throw std::runtime_error(
-            "custom allreduce only supports num gpus in (2,4,6,8). Actual num "
+            "custom allreduce only supports num gpus in (2,4,6,8). Actual "
+            "num "
             "gpus = " +
             std::to_string(world_size_));
     }
@@ -511,10 +595,11 @@ class CustomAllreduce {
     }
   }
 };
+
 /**
- * To inspect PTX/SASS, copy paste this header file to compiler explorer and add
- a template instantiation:
+ * To inspect PTX/SASS, copy paste this header file to compiler explorer and
+ add a template instantiation:
  * template void vllm::CustomAllreduce::allreduce<half>(cudaStream_t, half *,
  half *, int, int, int);
 */
-}  // namespace vllm
+}  // namespace vllm

csrc/custom_all_reduce_test.cu

@@ -1,9 +1,9 @@
 /**
  * This is a standalone test for custom allreduce.
  * To compile, make sure you have MPI and NCCL installed in your system.
- * export MPI_HOME=xxx
+ * export MPI_HOME=XXX
  * nvcc -O2 -arch=native -std=c++17 custom_all_reduce_test.cu -o
- * custom_all_reduce_test -lnccl -I${MPI_HOME} -lmpi
+ * custom_all_reduce_test -lnccl -I${MPI_HOME}/include -lmpi
  *
  * Warning: this C++ test is not designed to be very readable and was used
  * during the rapid prototyping process.
@@ -22,7 +22,15 @@
 #include "cuda_profiler_api.h"
 #include "custom_all_reduce.cuh"
 #include "mpi.h"
-#include "nccl.h"
+#ifdef USE_ROCM
+  #include <hip/hip_bf16.h>
+typedef __hip_bfloat16 nv_bfloat16;
+  #include "rccl/rccl.h"
+  #include "custom_all_reduce_hip.cuh"
+#else
+  #include "nccl.h"
+  #include "custom_all_reduce.cuh"
+#endif
 
 #define MPICHECK(cmd)                                                  \
   do {                                                                 \
@@ -43,16 +51,29 @@
     }                                                               \
   } while (0)
 
+#ifdef USE_ROCM
 __global__ void dummy_kernel() {
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
+  for (int i = 0; i < 100; i++) {
+    uint64_t start = wall_clock64();
+    uint64_t cycles_elapsed;
+    do {
+      cycles_elapsed = wall_clock64() - start;
+    } while (cycles_elapsed < 100);
+  }
   for (int i = 0; i < 100; i++) __nanosleep(1000000);  // 100ms
+}
 #else
+__global__ void dummy_kernel() {
+  #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
+  for (int i = 0; i < 100; i++) __nanosleep(1000000);  // 100ms
+  #else
   for (int i = 0; i < 100; i++) {
     long long int start = clock64();
     while (clock64() - start < 150000000);  // approximately 98.4ms on P40
   }
-#endif
+  #endif
 }
+#endif
 
 template <typename T>
 __global__ void set_data(T* data, int size, int myRank) {
@@ -121,8 +142,14 @@ void run(int myRank, int nRanks, ncclComm_t& comm, int threads, int block_limit,
    * registration, they are allocated and registered together in the test for
    * convenience.
    */
+#ifdef USE_ROCM
+  CUDACHECK(hipExtMallocWithFlags(
+      (void**)&buffer, 2 * data_size * sizeof(T) + sizeof(vllm::Signal),
+      hipDeviceMallocUncached));
+#else
   CUDACHECK(
       cudaMalloc(&buffer, 2 * data_size * sizeof(T) + sizeof(vllm::Signal)));
+#endif
   CUDACHECK(
       cudaMemset(buffer, 0, 2 * data_size * sizeof(T) + sizeof(vllm::Signal)));
   CUDACHECK(cudaMalloc(&self_data_copy, data_size * sizeof(T)));
@@ -311,13 +338,18 @@ int main(int argc, char** argv) {
 
   bool performance_test = true;
   cudaProfilerStart();
-  // Uncomment to scan through different block size configs.
-  // for (int threads : {256, 512, 1024}) {
-  //   for (int block_limit = 16; block_limit < 112; block_limit += 4) {
-  //     run<half>(myRank, nRanks, comm, threads, block_limit, 1024 * 1024,
-  //     performance_test);
-  //   }
-  // }
+// Uncomment to scan through different block size configs.
+// for (int threads : {256, 512, 1024}) {
+//   for (int block_limit = 16; block_limit < 112; block_limit += 4) {
+//     run<half>(myRank, nRanks, comm, threads, block_limit, 1024 * 1024,
+//     performance_test);
+//   }
+// }
+#ifdef USE_ROCM
+  const int block_limit = 16;
+#else
+  const int block_limit = 36;
+#endif
   // Scan through different sizes to test performance.
   for (int sz = 512; sz <= (8 << 20); sz *= 2) {
     run<half>(myRank, nRanks, comm, 512, 36, sz + 8 * 47, performance_test);
@@ -326,4 +358,4 @@ int main(int argc, char** argv) {
   cudaProfilerStop();
   MPICHECK(MPI_Finalize());
   return EXIT_SUCCESS;
-}
+}

csrc/cutlass_extensions/common.hpp

@@ -48,4 +48,14 @@ struct enable_sm90_or_later : Kernel {
     Kernel::operator()(std::forward<Args>(args)...);
 #endif
   }
-};
+};
+
+template <typename Kernel>
+struct enable_sm90_only : Kernel {
+  template <typename... Args>
+  CUTLASS_DEVICE void operator()(Args&&... args) {
+#if defined __CUDA_ARCH__ && __CUDA_ARCH__ == 900
+    Kernel::operator()(std::forward<Args>(args)...);
+#endif
+  }
+};

csrc/cutlass_extensions/epilogue/broadcast_load_epilogue_array_c3x.hpp

@@ -0,0 +1,457 @@
+/***************************************************************************************************
+ * Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights
+ *reserved. SPDX-License-Identifier: BSD-3-Clause
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are met:
+ *
+ * 1. Redistributions of source code must retain the above copyright notice,
+ *this list of conditions and the following disclaimer.
+ *
+ * 2. Redistributions in binary form must reproduce the above copyright notice,
+ * this list of conditions and the following disclaimer in the documentation
+ * and/or other materials provided with the distribution.
+ *
+ * 3. Neither the name of the copyright holder nor the names of its
+ * contributors may be used to endorse or promote products derived from
+ * this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+ * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ *ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
+ *LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+ *CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+ *SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+ *INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+ *CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ *ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+ *POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+
+//
+// This file is a modified excerpt of
+// include/cutlass/epilogue/fusion/sm90_visitor_load_tma_warpspecialized.hpp
+// from https://github.com/NVIDIA/cutlass v3.5.0
+// It has been modified to support either row/column or scalar broadcasting
+// where the tensor being loaded from is always passed in via a device pointer.
+// This lets one compiled kernel handle all cases of per-tensor or
+// per-channel/per-token quantization.
+//
+// This interface also allows the scales to be passed in as tensors that
+// consistently reside on the device, which avoids an issue with a previous
+// implementation where scalars needed to be on the CPU since they
+// were passed in via float values. This created a potential performance hazard
+// if scales were initially on the device, and caused torch.compile graphs
+// breaks when moving scales to the CPU.
+//
+#pragma once
+
+// Turn off clang-format for the entire file to keep it close to upstream
+// clang-format off
+
+#include "cutlass/cutlass.h"
+#include "cutlass/arch/barrier.h"
+
+#include "cute/tensor.hpp"
+#include "cutlass/epilogue/fusion/sm90_visitor_tma_warpspecialized.hpp"
+
+namespace cutlass::epilogue::fusion {
+
+using namespace cute;
+using namespace detail;
+
+// Row vector broadcast
+template<
+  int Stages,
+  class CtaTileShapeMNK,
+  class Element,
+  class StrideMNL = Stride<_0,_1,_0>,
+  int Alignment = 128 / sizeof_bits_v<Element>
+>
+struct Sm90RowOrScalarBroadcastArray {
+  static_assert(Stages == 0, "Row broadcast doesn't support smem usage");
+  static_assert(is_static_v<decltype(take<0,2>(StrideMNL{}))>); // batch stride can be dynamic or static
+  static_assert(take<0,2>(StrideMNL{}) == Stride<_0,_1>{});
+
+  struct SharedStorage { 
+    array_aligned<Element, size<1>(CtaTileShapeMNK{})> smem;
+  };
+
+  // This struct has been modified to have a bool indicating that ptr_row is a 
+  // scalar that must be broadcast, instead of containing a scalar that is 
+  // valid if ptr_row is null.
+  struct Arguments {
+    const Element* const* ptr_row_array = nullptr;
+    bool row_broadcast = true;
+    StrideMNL dRow = {};
+  };
+
+  using Params = Arguments;
+
+  template <class ProblemShape>
+  static constexpr Params
+  to_underlying_arguments(ProblemShape const& problem_shape, Arguments const& args, void* workspace) {
+    return args;
+  }
+
+  template <class ProblemShape>
+  static bool
+  can_implement(ProblemShape const& problem_shape, Arguments const& args) {
+    return true;
+  }
+
+  template <class ProblemShape>
+  static size_t
+  get_workspace_size(ProblemShape const& problem_shape, Arguments const& args) {
+    return 0;
+  }
+
+  template <class ProblemShape>
+  static cutlass::Status
+  initialize_workspace(ProblemShape const& problem_shape, Arguments const& args, void* workspace, cudaStream_t stream,
+    CudaHostAdapter* cuda_adapter = nullptr) {
+    return cutlass::Status::kSuccess;
+  }
+
+  CUTLASS_HOST_DEVICE
+  Sm90RowOrScalarBroadcastArray() { }
+
+  CUTLASS_HOST_DEVICE
+  Sm90RowOrScalarBroadcastArray(Params const& params, SharedStorage const& shared_storage)
+      : params(params)
+      , smem(const_cast<Element*>(shared_storage.smem.data())) { }
+
+  Params params;
+  Element *smem = nullptr;
+
+  CUTLASS_DEVICE bool
+  is_producer_load_needed() const {
+    return false;
+  }
+
+  CUTLASS_DEVICE bool
+  is_C_load_needed() const {
+    return false;
+  }
+
+  CUTLASS_DEVICE bool
+  is_zero() const {
+    return (!params.row_broadcast && *(params.ptr_row_array[group]) == Element(0));
+  }
+
+  template <class... Args>
+  CUTLASS_DEVICE auto
+  get_producer_load_callbacks(ProducerLoadArgs<Args...> const& args) {
+    return EmptyProducerLoadCallbacks{};
+  }
+
+  template <class GS_GTensor, class GS_STensor, class GS_CTensor, class Tiled_G2S, class SR_STensor, class SR_RTensor, class CTensor, class ThrResidue, class ThrNum>
+  struct ConsumerStoreCallbacks : EmptyConsumerStoreCallbacks {
+    CUTLASS_DEVICE
+    ConsumerStoreCallbacks(
+        GS_GTensor tGS_gRow_, GS_STensor tGS_sRow_, 
+        GS_CTensor tGS_cRow_, Tiled_G2S tiled_g2s_, 
+        SR_STensor tSR_sRow_, SR_RTensor tSR_rRow_,
+        CTensor tCcRow_, ThrResidue residue_tCcRow_, ThrNum thr_num_,
+        int group, Params const& params_)
+      : tGS_gRow(tGS_gRow_)
+      , tGS_sRow(tGS_sRow_)
+      , tGS_cRow(tGS_cRow_)
+      , tiled_G2S(tiled_g2s_)
+      , tSR_sRow(tSR_sRow_)
+      , tSR_rRow(tSR_rRow_)
+      , tCcRow(tCcRow_)
+      , residue_tCcRow(residue_tCcRow_)
+      , group(group)
+      , params(params_) {}
+
+    GS_GTensor tGS_gRow;                                                         // (CPY,CPY_M,CPY_N)
+    GS_STensor tGS_sRow;                                                         // (CPY,CPY_M,CPY_N)
+    GS_CTensor tGS_cRow;                                                         // (CPY,CPY_M,CPY_N)
+    Tiled_G2S tiled_G2S;
+
+    SR_STensor tSR_sRow;                                                         // (CPY,CPY_M,CPY_N,EPI_M,EPI_N)
+    SR_RTensor tSR_rRow;                                                         // (CPY,CPY_M,CPY_N,EPI_M,EPI_N) 
+  
+    CTensor tCcRow;                                                              // (CPY,CPY_M,CPY_N,EPI_M,EPI_N)
+    ThrResidue residue_tCcRow;                                                   // (m, n)
+    ThrNum thr_num;
+    int group;
+    Params const& params;
+
+    CUTLASS_DEVICE void
+    begin() {
+      if (!params.row_broadcast) {
+        fill(tSR_rRow, *(params.ptr_row_array[group]));
+        return;
+      }
+
+      auto synchronize = [&] () { cutlass::arch::NamedBarrier::sync(thr_num, cutlass::arch::ReservedNamedBarriers::EpilogueBarrier); };
+      Tensor tGS_gRow_flt = filter_zeros(tGS_gRow);
+      Tensor tGS_sRow_flt = filter_zeros(tGS_sRow);
+      Tensor tGS_cRow_flt = make_tensor(tGS_cRow.data(), make_layout(tGS_gRow_flt.shape(), tGS_cRow.stride()));
+
+      for (int i = 0; i < size(tGS_gRow_flt); ++i) {
+        if (get<1>(tGS_cRow_flt(i)) >= size<1>(CtaTileShapeMNK{})) {
+          continue; // OOB of SMEM, 
+        }
+        if (elem_less(tGS_cRow_flt(i), make_coord(get<0>(residue_tCcRow), get<1>(residue_tCcRow)))) {
+          tGS_sRow_flt(i) = tGS_gRow_flt(i);
+        }
+        else {
+          tGS_sRow_flt(i) = Element(0); // Set to Zero when OOB so LDS could be issue without any preds.
+        }
+      }
+      synchronize();
+    }
+
+    CUTLASS_DEVICE void
+    begin_loop(int epi_m, int epi_n) {
+      if (epi_m == 0) { // Assumes M-major subtile loop
+        if (!params.row_broadcast) return; // Do not issue LDS when row is scalar 
+        Tensor tSR_sRow_flt = filter_zeros(tSR_sRow(_,_,_,epi_m,epi_n));
+        Tensor tSR_rRow_flt = filter_zeros(tSR_rRow);
+        copy(tSR_sRow_flt, tSR_rRow_flt);
+      }
+    }
+
+    template <typename ElementAccumulator, int FragmentSize>
+    CUTLASS_DEVICE Array<Element, FragmentSize>
+    visit(Array<ElementAccumulator, FragmentSize> const& frg_acc, int epi_v, int epi_m, int epi_n) {
+      Array<Element, FragmentSize> frg_row;
+
+      CUTLASS_PRAGMA_UNROLL
+      for (int i = 0; i < FragmentSize; ++i) {
+        frg_row[i] = tSR_rRow(epi_v * FragmentSize + i);
+      }
+
+      return frg_row;
+    }
+  };
+
+  template <
+    bool ReferenceSrc, // do register tensors reference the src or dst layout of the tiled copy
+    class... Args
+  >
+  CUTLASS_DEVICE auto
+  get_consumer_store_callbacks(ConsumerStoreArgs<Args...> const& args) {
+    auto [M, N, K, L] = args.problem_shape_mnkl;
+    auto [m, n, k, l] = args.tile_coord_mnkl;
+    using ThreadCount = decltype(size(args.tiled_copy));
+
+    Tensor mRow = make_tensor(make_gmem_ptr(params.ptr_row_array[l]), make_shape(M,N,1), params.dRow);
+    Tensor gRow = local_tile(mRow(_,_,l), take<0,2>(args.tile_shape_mnk), make_coord(m, n));          // (CTA_M, CTA_N)
+    Tensor sRow = make_tensor(make_smem_ptr(smem), 
+        make_shape(size<0>(CtaTileShapeMNK{}), size<1>(CtaTileShapeMNK{})), make_shape(_0{}, _1{}));  // (CTA_M, CTA_N)
+    //// G2S: Gmem to Smem
+    auto tiled_g2s = make_tiled_copy(Copy_Atom<DefaultCopy, Element>{},
+                                     Layout< Shape<_1, ThreadCount>, 
+                                            Stride<_0,          _1>>{}, 
+                                     Layout<_1>{});   
+    auto thr_g2s = tiled_g2s.get_slice(args.thread_idx);
+    Tensor tGS_gRow = thr_g2s.partition_S(gRow);
+    Tensor tGS_sRow = thr_g2s.partition_D(sRow);
+
+    //// G2S: Coord 
+    auto cRow = make_identity_tensor(make_shape(size<0>(CtaTileShapeMNK{}), size<1>(CtaTileShapeMNK{})));
+    Tensor tGS_cRow = thr_g2s.partition_S(cRow);
+
+    //// S2R: Smem to Reg
+    Tensor tSR_sRow = sm90_partition_for_epilogue<ReferenceSrc>(sRow, args.epi_tile, args.tiled_copy, args.thread_idx);
+    Tensor tSR_rRow = make_tensor_like(take<0,3>(tSR_sRow));                                           // (CPY,CPY_M,CPY_N)
+
+    return ConsumerStoreCallbacks<decltype(tGS_gRow), decltype(tGS_sRow), decltype(tGS_cRow), decltype(tiled_g2s), decltype(tSR_sRow), decltype(tSR_rRow), decltype(args.tCcD), decltype(args.residue_cD), ThreadCount>(
+      tGS_gRow, 
+      tGS_sRow, 
+      tGS_cRow, tiled_g2s, 
+      tSR_sRow, 
+      tSR_rRow, 
+      args.tCcD, 
+      args.residue_cD,
+      ThreadCount{}, 
+      l,
+      params);
+  }
+};
+
+/////////////////////////////////////////////////////////////////////////////////////////////////
+
+// Column vector broadcast
+template<
+  int Stages,
+  class CtaTileShapeMNK,
+  class Element,
+  class StrideMNL = Stride<_1,_0,_0>,
+  int Alignment = 128 / sizeof_bits_v<Element>
+>
+struct Sm90ColOrScalarBroadcastArray {
+  static_assert(Stages == 0, "Column broadcast doesn't support smem usage yet");
+  static_assert(Alignment * sizeof_bits_v<Element> % 128 == 0, "sub-16B alignment not supported yet");
+  static_assert(
+    (cute::is_same_v<StrideMNL, Stride<_1,_0, _0>>) || // col vector broadcast, e.g. per-row alpha/bias
+    (cute::is_same_v<StrideMNL, Stride<_1,_0,int>>));  // batched col vector broadcast, e.g. batched per-row bias
+
+  // Accumulator distributes col elements evenly amongst threads so we can just directly load from gmem
+  struct SharedStorage { };
+
+  // This struct has been modified to have a bool indicating that ptr_col is a 
+  // scalar that must be broadcast, instead of containing a scalar that is 
+  // valid if ptr_col is null.
+  struct Arguments {
+    const Element* const* ptr_col_array = nullptr;
+    bool col_broadcast = true;
+    StrideMNL dCol = {};
+  };
+
+  using Params = Arguments;
+
+  template <class ProblemShape>
+  static constexpr Params
+  to_underlying_arguments(ProblemShape const& problem_shape, Arguments const& args, void* workspace) {
+    return args;
+  }
+
+  template <class ProblemShape>
+  static bool
+  can_implement(ProblemShape const& problem_shape, Arguments const& args) {
+    return true;
+  }
+
+  template <class ProblemShape>
+  static size_t
+  get_workspace_size(ProblemShape const& problem_shape, Arguments const& args) {
+    return 0;
+  }
+
+  template <class ProblemShape>
+  static cutlass::Status
+  initialize_workspace(ProblemShape const& problem_shape, Arguments const& args, void* workspace, cudaStream_t stream,
+    CudaHostAdapter* cuda_adapter = nullptr) {
+    return cutlass::Status::kSuccess;
+  }
+
+  CUTLASS_DEVICE bool
+  is_producer_load_needed() const {
+    return false;
+  }
+
+  CUTLASS_DEVICE bool
+  is_C_load_needed() const {
+    return false;
+  }
+
+  CUTLASS_DEVICE bool
+  is_zero() const {
+    return (!params.col_broadcast && *(params.ptr_col_array[group]) == Element(0));
+  }
+
+  CUTLASS_HOST_DEVICE
+  Sm90ColOrScalarBroadcastArray() { }
+
+  CUTLASS_HOST_DEVICE
+  Sm90ColOrScalarBroadcastArray(Params const& params, SharedStorage const& shared_storage)
+      : params(params) { }
+
+  Params params;
+
+  template <class... Args>
+  CUTLASS_DEVICE auto
+  get_producer_load_callbacks(ProducerLoadArgs<Args...> const& args) {
+    return EmptyProducerLoadCallbacks{};
+  }
+
+  template<class GTensor, class RTensor, class CTensor, class ProblemShape>
+  struct ConsumerStoreCallbacks : EmptyConsumerStoreCallbacks {
+    CUTLASS_DEVICE
+    ConsumerStoreCallbacks(
+      GTensor&& tCgCol,
+      RTensor&& tCrCol,
+      CTensor&& tCcCol,
+      ProblemShape problem_shape,
+      int group,
+      Params const& params
+    ): 
+      tCgCol(cute::forward<GTensor>(tCgCol)),
+      tCrCol(cute::forward<RTensor>(tCrCol)),
+      tCcCol(cute::forward<CTensor>(tCcCol)),
+      m(get<0>(problem_shape)),
+      group(group),
+      params(params) {}
+
+    GTensor tCgCol;                                                                    // (CPY,CPY_M,CPY_N,EPI_M,EPI_N)
+    RTensor tCrCol;
+    CTensor tCcCol;                                                                    // (CPY,CPY_M,CPY_N,EPI_M,EPI_N)
+    Params const& params;
+    int m;
+    int group;
+
+    CUTLASS_DEVICE void
+    begin() {
+      Tensor pred = make_tensor<bool>(shape(tCgCol));
+      CUTLASS_PRAGMA_UNROLL
+      for (int i = 0; i < size(pred); ++i) {
+        pred(i) = get<0>(tCcCol(i)) < m;
+      }
+
+      if (!params.col_broadcast) {
+        fill(tCrCol, *(params.ptr_col_array[group]));
+        return;
+      }
+
+      // Filter so we don't issue redundant copies over stride-0 modes
+      // (only works if 0-strides are in same location, which is by construction)
+      copy_if(pred, filter(tCgCol), filter(tCrCol));
+    }
+
+    template <typename ElementAccumulator, int FragmentSize>
+    CUTLASS_DEVICE Array<Element, FragmentSize>
+    visit(Array<ElementAccumulator, FragmentSize> const& frg_acc, int epi_v, int epi_m, int epi_n) {
+      Array<Element, FragmentSize> frg_col;
+      Tensor tCrCol_mn = tCrCol(_,_,_,epi_m,epi_n);
+
+      CUTLASS_PRAGMA_UNROLL
+      for (int i = 0; i < FragmentSize; ++i) {
+        frg_col[i] = tCrCol_mn(epi_v * FragmentSize + i);
+      }
+
+      return frg_col;
+    }
+
+  };
+
+  template <
+    bool ReferenceSrc, // do register tensors reference the src or dst layout of the tiled copy
+    class... Args
+  >
+  CUTLASS_DEVICE auto
+  get_consumer_store_callbacks(ConsumerStoreArgs<Args...> const& args) {
+
+    auto [M, N, K, L] = args.problem_shape_mnkl;
+    auto [m, n, k, l] = args.tile_coord_mnkl;
+
+    Tensor mCol = make_tensor(make_gmem_ptr(params.ptr_col_array[l]), make_shape(M,N,1), params.dCol);
+    Tensor tCgCol = sm90_partition_for_epilogue<ReferenceSrc>(                         // (CPY,CPY_M,CPY_N,EPI_M,EPI_N)
+      mCol, args.tile_shape_mnk, args.tile_coord_mnkl, args.epi_tile, args.tiled_copy, args.thread_idx);
+    Tensor tCrCol = make_tensor_like(tCgCol);                                          // (CPY,CPY_M,CPY_N,EPI_M,EPI_N)
+
+    // Generate an identity tensor matching the shape of the global tensor and 
+    //  partition the same way, this will be used to generate the predicate
+    //  tensor for loading
+    Tensor cCol = make_identity_tensor(mCol.shape());
+    Tensor tCcCol = sm90_partition_for_epilogue<ReferenceSrc>(                         // (CPY,CPY_M,CPY_N,EPI_M,EPI_N)
+      cCol, args.tile_shape_mnk, args.tile_coord_mnkl, args.epi_tile, args.tiled_copy, args.thread_idx);
+
+    return ConsumerStoreCallbacks(
+      cute::move(tCgCol), 
+      cute::move(tCrCol), 
+      cute::move(tCcCol), 
+      args.problem_shape_mnkl,
+      l,
+      params
+    );
+  }
+};
+
+}

csrc/cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp

@@ -1,6 +1,7 @@
 #pragma once
 
 #include "cutlass_extensions/epilogue/broadcast_load_epilogue_c3x.hpp"
+#include "cutlass_extensions/epilogue/broadcast_load_epilogue_array_c3x.hpp"
 
 /*
    This file defines custom epilogues for fusing channel scales, token scales,
@@ -69,6 +70,16 @@ struct ScaledEpilogueBase {
       0 /*Stages*/, TileShape, T, T, Stride<Int<0>, Int<1>, Int<0>>,
       128 / sizeof_bits_v<T>, EnableNullPtr>;
 
+  template <typename T>
+  using ColOrScalarLoadArray =
+      cutlass::epilogue::fusion::Sm90ColOrScalarBroadcastArray<
+          0 /*Stages*/, TileShape, T, Stride<Int<1>, Int<0>, Int<0>>>;
+
+  template <typename T>
+  using RowOrScalarLoadArray =
+      cutlass::epilogue::fusion::Sm90RowOrScalarBroadcastArray<
+          0 /*Stages*/, TileShape, T, Stride<Int<0>, Int<1>, Int<0>>>;
+
   // This utility function constructs the arguments for the load descriptors
   // from a tensor. It can handle both row and column, as well as row/column or
   // scalar cases.
@@ -96,6 +107,14 @@ struct ScaledEpilogueBase {
                   std::is_same_v<Descriptor, RowLoad<T, true>>);
     return Arguments{data_ptr};
   }
+
+  template <typename Descriptor, typename T>
+  static auto args_from_tensor(const T* const* data_ptr, bool do_broadcast) {
+    using Arguments = typename Descriptor::Arguments;
+    static_assert(std::is_same_v<Descriptor, ColOrScalarLoadArray<T>> ||
+                  std::is_same_v<Descriptor, RowOrScalarLoadArray<T>>);
+    return Arguments{data_ptr, do_broadcast};
+  }
 };
 
 /*
@@ -381,4 +400,51 @@ struct ScaledEpilogueBiasAzpToken
   }
 };
 
+/*
+    This epilogue works like ScaledEpilogue, but ScaleA and ScaleB are pointers
+    to arrays containing different scales used in group gemm. The number of
+   pointers in ScaleA and the number of pointers in ScaleB are equal to the
+   group size.
+*/
+template <typename ElementAcc, typename ElementD, typename EpilogueDescriptor>
+struct ScaledEpilogueArray
+    : private ScaledEpilogueBase<ElementAcc, ElementD, EpilogueDescriptor> {
+ private:
+  using SUPER = ScaledEpilogueBase<ElementAcc, ElementD, EpilogueDescriptor>;
+  using Accum = typename SUPER::Accum;
+  using ScaleA = typename SUPER::template ColOrScalarLoadArray<float>;
+  using ScaleB = typename SUPER::template RowOrScalarLoadArray<float>;
+
+  using Compute0 = cutlass::epilogue::fusion::Sm90Compute<
+      cutlass::multiplies, float, float,
+      cutlass::FloatRoundStyle::round_to_nearest>;
+
+  using EVTCompute0 =
+      cutlass::epilogue::fusion::Sm90EVT<Compute0, ScaleB, Accum>;
+
+  using Compute1 = cutlass::epilogue::fusion::Sm90Compute<
+      cutlass::multiplies, ElementD, float,
+      cutlass::FloatRoundStyle::round_to_nearest>;
+
+ public:
+  using EVTCompute =
+      cutlass::epilogue::fusion::Sm90EVT<Compute1, ScaleA, EVTCompute0>;
+  using ArgumentType = typename EVTCompute::Arguments;
+
+  using ScaleAArray = typename SUPER::template ColOrScalarLoadArray<float>;
+  using ScaleBArray = typename SUPER::template RowOrScalarLoadArray<float>;
+
+  static ArgumentType prepare_args(float const* const* a_scales_ptr,
+                                   float const* const* b_scales_ptr,
+                                   bool a_col_broadcast, bool b_row_broadcast) {
+    auto a_args = SUPER::template args_from_tensor<ScaleAArray, float>(
+        a_scales_ptr, a_col_broadcast);
+    auto b_args = SUPER::template args_from_tensor<ScaleBArray, float>(
+        b_scales_ptr, b_row_broadcast);
+
+    typename EVTCompute0::Arguments evt0_args{b_args, {}, {}};
+    return ArgumentType{a_args, evt0_args, {}};
+  }
+};
+
 };  // namespace vllm::c3x

csrc/ops.h

@@ -119,6 +119,8 @@ void advance_step_flashinfer(
     torch::Tensor& paged_kv_indices, torch::Tensor& paged_kv_indptr,
     torch::Tensor& paged_kv_last_page_len, torch::Tensor& block_table_bounds);
 
+torch::Tensor get_cuda_view_from_cpu_tensor(torch::Tensor& cpu_tensor);
+
 #ifndef USE_ROCM
 torch::Tensor aqlm_gemm(const torch::Tensor& input, const torch::Tensor& codes,
                         const torch::Tensor& codebooks,
@@ -143,7 +145,8 @@ torch::Tensor permute_cols(torch::Tensor const& A, torch::Tensor const& perm);
 #endif
 
 torch::Tensor ggml_dequantize(torch::Tensor W, int64_t type, int64_t m,
-                              int64_t n);
+                              int64_t n,
+                              std::optional<at::ScalarType> const& dtype);
 
 torch::Tensor ggml_mul_mat_vec_a8(torch::Tensor W, torch::Tensor X,
                                   int64_t type, int64_t row);
@@ -164,6 +167,7 @@ int64_t ggml_moe_get_block_size(int64_t type);
 bool cutlass_scaled_mm_supports_fp4(int64_t cuda_device_capability);
 bool cutlass_scaled_mm_supports_fp8(int64_t cuda_device_capability);
 bool cutlass_scaled_mm_supports_block_fp8(int64_t cuda_device_capability);
+bool cutlass_group_gemm_supported(int64_t cuda_device_capability);
 
 void cutlass_scaled_fp4_mm(torch::Tensor& D, torch::Tensor const& A,
                            torch::Tensor const& B, torch::Tensor const& A_sf,
@@ -175,6 +179,19 @@ void cutlass_scaled_mm(torch::Tensor& out, torch::Tensor const& a,
                        torch::Tensor const& b_scales,
                        std::optional<torch::Tensor> const& bias);
 
+void cutlass_moe_mm(
+    torch::Tensor& out_tensors, torch::Tensor const& a_tensors,
+    torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
+    torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
+    torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
+    torch::Tensor const& b_strides, torch::Tensor const& c_strides);
+
+void get_cutlass_moe_mm_data(
+    const torch::Tensor& topk_ids, torch::Tensor& expert_offsets,
+    torch::Tensor& problem_sizes1, torch::Tensor& problem_sizes2,
+    torch::Tensor& input_permutation, torch::Tensor& output_permutation,
+    const int64_t num_experts, const int64_t n, const int64_t k);
+
 void cutlass_scaled_mm_azp(torch::Tensor& out, torch::Tensor const& a,
                            torch::Tensor const& b,
                            torch::Tensor const& a_scales,
@@ -251,10 +268,10 @@ void causal_conv1d_fwd(const at::Tensor& x, const at::Tensor& weight,
                        const std::optional<at::Tensor>& has_initial_state,
                        bool silu_activation, int64_t pad_slot_id);
 
-#ifndef USE_ROCM
 using fptr_t = int64_t;
 fptr_t init_custom_ar(const std::vector<int64_t>& fake_ipc_ptrs,
-                      torch::Tensor& rank_data, int64_t rank, bool full_nvlink);
+                      torch::Tensor& rank_data, int64_t rank,
+                      bool fully_connected);
 void all_reduce(fptr_t _fa, torch::Tensor& inp, torch::Tensor& out,
                 fptr_t reg_buffer, int64_t reg_buffer_sz_bytes);
 void dispose(fptr_t _fa);
@@ -265,4 +282,7 @@ get_graph_buffer_ipc_meta(fptr_t _fa);
 void register_graph_buffers(fptr_t _fa,
                             const std::vector<std::vector<int64_t>>& handles,
                             const std::vector<std::vector<int64_t>>& offsets);
-#endif
+std::tuple<int64_t, torch::Tensor> allocate_shared_buffer_and_handle(
+    int64_t size);
+int64_t open_mem_handle(torch::Tensor& mem_handle);
+void free_shared_buffer(int64_t buffer);

csrc/quantization/cutlass_w8a8/moe/get_group_starts.cuh

@@ -0,0 +1,80 @@
+#pragma once
+
+#include <cuda.h>
+#include <torch/all.h>
+#include <c10/cuda/CUDAStream.h>
+
+#include "core/scalar_type.hpp"
+#include "cutlass/bfloat16.h"
+#include "cutlass/float8.h"
+
+template <typename ElementAB, typename ElementC, typename ElementAccumulator>
+__global__ void get_group_gemm_starts(
+    int32_t* expert_offsets, ElementAB** a_offsets, ElementAB** b_offsets,
+    ElementC** out_offsets, ElementAccumulator** a_scales_offsets,
+    ElementAccumulator** b_scales_offsets, ElementAB* a_base_as_int,
+    ElementAB* b_base_as_int, ElementC* out_base_as_int,
+    ElementAccumulator* a_scales_base_as_int,
+    ElementAccumulator* b_scales_base_as_int, int64_t n, int64_t k,
+    bool per_act_token, bool per_out_ch) {
+  int expert_id = threadIdx.x;
+
+  int64_t expert_offset = expert_offsets[expert_id];
+
+  a_offsets[expert_id] = a_base_as_int + expert_offset * k;
+  b_offsets[expert_id] = b_base_as_int + expert_id * k * n;
+  out_offsets[expert_id] = out_base_as_int + expert_offset * n;
+  a_scales_offsets[expert_id] =
+      a_scales_base_as_int + (per_act_token ? expert_offset : 0);
+  b_scales_offsets[expert_id] =
+      b_scales_base_as_int + (per_out_ch ? n * expert_id : expert_id);
+}
+
+#define __CALL_GET_STARTS_KERNEL(TENSOR_C_TYPE, C_TYPE)                    \
+  else if (out_tensors.dtype() == TENSOR_C_TYPE) {                         \
+    get_group_gemm_starts<cutlass::float_e4m3_t, C_TYPE, float>            \
+        <<<1, num_experts, 0, stream>>>(                                   \
+            static_cast<int32_t*>(expert_offsets.data_ptr()),              \
+            static_cast<cutlass::float_e4m3_t**>(a_ptrs.data_ptr()),       \
+            static_cast<cutlass::float_e4m3_t**>(b_ptrs.data_ptr()),       \
+            static_cast<C_TYPE**>(out_ptrs.data_ptr()),                    \
+            static_cast<float**>(a_scales_ptrs.data_ptr()),                \
+            static_cast<float**>(b_scales_ptrs.data_ptr()),                \
+            static_cast<cutlass::float_e4m3_t*>(a_tensors.data_ptr()),     \
+            static_cast<cutlass::float_e4m3_t*>(b_tensors.data_ptr()),     \
+            static_cast<C_TYPE*>(out_tensors.data_ptr()),                  \
+            static_cast<float*>(a_scales.data_ptr()),                      \
+            static_cast<float*>(b_scales.data_ptr()), out_tensors.size(1), \
+            a_tensors.size(1), per_act_token, per_out_ch);                 \
+  }
+
+namespace {
+
+void run_get_group_gemm_starts(
+    torch::Tensor const& expert_offsets, torch::Tensor& a_ptrs,
+    torch::Tensor& b_ptrs, torch::Tensor& out_ptrs,
+    torch::Tensor& a_scales_ptrs, torch::Tensor& b_scales_ptrs,
+    torch::Tensor const& a_tensors, torch::Tensor const& b_tensors,
+    torch::Tensor& out_tensors, torch::Tensor const& a_scales,
+    torch::Tensor const& b_scales) {
+  TORCH_CHECK(a_tensors.dtype() == torch::kFloat8_e4m3fn);
+  TORCH_CHECK(b_tensors.dtype() == torch::kFloat8_e4m3fn);
+  TORCH_CHECK(a_scales.dtype() == torch::kFloat32);
+  TORCH_CHECK(b_scales.dtype() == torch::kFloat32);
+
+  int num_experts = static_cast<int>(expert_offsets.size(0));
+  bool per_act_token = a_scales.numel() != 1;
+  bool per_out_ch = b_scales.numel() != num_experts;
+
+  auto stream = at::cuda::getCurrentCUDAStream(a_tensors.device().index());
+
+  if (false) {
+  }
+  __CALL_GET_STARTS_KERNEL(torch::kBFloat16, cutlass::bfloat16_t)
+  __CALL_GET_STARTS_KERNEL(torch::kFloat16, half)
+  else {
+    TORCH_CHECK(false, "Invalid output type (must be float16 or bfloat16)");
+  }
+}
+
+}  // namespace

csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu

@@ -0,0 +1,160 @@
+#include <cudaTypedefs.h>
+
+#include <c10/cuda/CUDAGuard.h>
+#include <torch/all.h>
+
+#include "cutlass/cutlass.h"
+#include "grouped_mm_c3x.cuh"
+
+using namespace cute;
+
+namespace {
+
+template <typename InType, typename OutType,
+          template <typename, typename, typename> typename Epilogue>
+struct sm90_fp8_config_default {
+  // M in (16, inf)
+  static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
+  using KernelSchedule =
+      cutlass::gemm::KernelPtrArrayTmaWarpSpecializedPingpongFP8FastAccum;
+  using EpilogueSchedule =
+      cutlass::epilogue::PtrArrayTmaWarpSpecializedPingpong;
+  using TileShape = cute::Shape<cute::_64, cute::_256, cute::_128>;
+  using ClusterShape = cute::Shape<cute::_1, cute::_2, cute::_1>;
+
+  using Cutlass3xGemm =
+      cutlass_3x_group_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
+                            KernelSchedule, EpilogueSchedule>;
+};
+
+template <typename InType, typename OutType,
+          template <typename, typename, typename> typename Epilogue>
+struct sm90_fp8_config_M16 {
+  // M in [1, 16]
+  static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
+  using KernelSchedule =
+      cutlass::gemm::KernelPtrArrayTmaWarpSpecializedPingpongFP8FastAccum;
+  using EpilogueSchedule =
+      cutlass::epilogue::PtrArrayTmaWarpSpecializedPingpong;
+  using TileShape = cute::Shape<cute::_64, cute::_64, cute::_128>;
+  using ClusterShape = cute::Shape<cute::_1, cute::_4, cute::_1>;
+
+  using Cutlass3xGemm =
+      cutlass_3x_group_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
+                            KernelSchedule, EpilogueSchedule>;
+};
+
+template <typename InType, typename OutType,
+          template <typename, typename, typename> typename Epilogue>
+struct sm90_fp8_config_K8192 {
+  // K in [8192, inf)
+  static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
+  using KernelSchedule =
+      cutlass::gemm::KernelPtrArrayTmaWarpSpecializedPingpongFP8FastAccum;
+  using EpilogueSchedule =
+      cutlass::epilogue::PtrArrayTmaWarpSpecializedPingpong;
+  using TileShape = cute::Shape<cute::_128, cute::_128, cute::_128>;
+  using ClusterShape = cute::Shape<cute::_1, cute::_8, cute::_1>;
+
+  using Cutlass3xGemm =
+      cutlass_3x_group_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
+                            KernelSchedule, EpilogueSchedule>;
+};
+
+template <typename InType, typename OutType,
+          template <typename, typename, typename> typename Epilogue>
+struct sm90_fp8_config_N8192 {
+  // N in [8192, inf)
+  static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
+  using KernelSchedule =
+      cutlass::gemm::KernelPtrArrayTmaWarpSpecializedPingpongFP8FastAccum;
+  using EpilogueSchedule =
+      cutlass::epilogue::PtrArrayTmaWarpSpecializedPingpong;
+  using TileShape = cute::Shape<cute::_64, cute::_128, cute::_256>;
+  using ClusterShape = cute::Shape<cute::_1, cute::_8, cute::_1>;
+
+  using Cutlass3xGemm =
+      cutlass_3x_group_gemm<InType, OutType, Epilogue, TileShape, ClusterShape,
+                            KernelSchedule, EpilogueSchedule>;
+};
+
+template <typename InType, typename OutType>
+void run_cutlass_moe_mm_sm90(
+    torch::Tensor& out_tensors, torch::Tensor const& a_tensors,
+    torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
+    torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
+    torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
+    torch::Tensor const& b_strides, torch::Tensor const& c_strides) {
+  TORCH_CHECK(a_tensors.size(0) > 0, "No input A tensors provided.");
+  TORCH_CHECK(b_tensors.size(0) > 0, "No input B tensors provided.");
+  TORCH_CHECK(out_tensors.size(0) > 0, "No output tensors provided.");
+
+  TORCH_CHECK(a_tensors.dtype() == torch::kFloat8_e4m3fn,
+              "A tensors must be of type float8_e4m3fn.");
+  TORCH_CHECK(b_tensors.dtype() == torch::kFloat8_e4m3fn,
+              "B tensors must be of type float8_e4m3fn.");
+
+  TORCH_CHECK(a_tensors.dtype() == torch::kFloat8_e4m3fn);
+  TORCH_CHECK(b_tensors.dtype() == torch::kFloat8_e4m3fn);
+
+  using Cutlass3xGemmN8192 = typename sm90_fp8_config_N8192<
+      InType, OutType, vllm::c3x::ScaledEpilogueArray>::Cutlass3xGemm;
+  using Cutlass3xGemmK8192 = typename sm90_fp8_config_K8192<
+      InType, OutType, vllm::c3x::ScaledEpilogueArray>::Cutlass3xGemm;
+  using Cutlass3xGemmM16 = typename sm90_fp8_config_M16<
+      InType, OutType, vllm::c3x::ScaledEpilogueArray>::Cutlass3xGemm;
+  using Cutlass3xGemmDefault = typename sm90_fp8_config_default<
+      InType, OutType, vllm::c3x::ScaledEpilogueArray>::Cutlass3xGemm;
+
+  uint32_t const m = a_tensors.size(0);
+  uint32_t const n = out_tensors.size(1);
+  uint32_t const k = a_tensors.size(1);
+
+  if (n >= 8192) {
+    cutlass_group_gemm_caller<Cutlass3xGemmN8192>(
+        out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
+        problem_sizes, a_strides, b_strides, c_strides);
+  } else if (k >= 8192) {
+    cutlass_group_gemm_caller<Cutlass3xGemmK8192>(
+        out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
+        problem_sizes, a_strides, b_strides, c_strides);
+  } else if (m <= 16) {
+    cutlass_group_gemm_caller<Cutlass3xGemmM16>(
+        out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
+        problem_sizes, a_strides, b_strides, c_strides);
+  } else {
+    cutlass_group_gemm_caller<Cutlass3xGemmDefault>(
+        out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
+        problem_sizes, a_strides, b_strides, c_strides);
+  }
+}
+
+void dispatch_moe_mm_sm90(
+    torch::Tensor& out_tensors, torch::Tensor const& a_tensors,
+    torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
+    torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
+    torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
+    torch::Tensor const& b_strides, torch::Tensor const& c_strides) {
+  if (out_tensors.dtype() == torch::kBFloat16) {
+    run_cutlass_moe_mm_sm90<cutlass::float_e4m3_t, cutlass::bfloat16_t>(
+        out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
+        problem_sizes, a_strides, b_strides, c_strides);
+  } else {
+    run_cutlass_moe_mm_sm90<cutlass::float_e4m3_t, cutlass::half_t>(
+        out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
+        problem_sizes, a_strides, b_strides, c_strides);
+  }
+}
+
+}  // namespace
+
+void cutlass_moe_mm_sm90(
+    torch::Tensor& out_tensors, torch::Tensor const& a_tensors,
+    torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
+    torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
+    torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
+    torch::Tensor const& b_strides, torch::Tensor const& c_strides) {
+  dispatch_moe_mm_sm90(out_tensors, a_tensors, b_tensors, a_scales, b_scales,
+                       expert_offsets, problem_sizes, a_strides, b_strides,
+                       c_strides);
+}

csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cuh

@@ -0,0 +1,149 @@
+#pragma once
+
+#include "cutlass/cutlass.h"
+
+#include "cutlass/gemm/collective/collective_builder.hpp"
+#include "cutlass/epilogue/collective/collective_builder.hpp"
+#include "cutlass/gemm/device/gemm_universal_adapter.h"
+
+#include "cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
+#include "cutlass_extensions/common.hpp"
+#include "get_group_starts.cuh"
+
+using namespace cute;
+
+namespace {
+
+using ProblemShape =
+    cutlass::gemm::GroupProblemShape<cute::Shape<int, int, int>>;
+
+using ElementAccumulator = float;
+using ArchTag = cutlass::arch::Sm90;
+using OperatorClass = cutlass::arch::OpClassTensorOp;
+
+using LayoutA = cutlass::layout::RowMajor;
+using LayoutB = cutlass::layout::ColumnMajor;
+using LayoutC = cutlass::layout::RowMajor;
+
+template <typename ElementAB_, typename ElementC_,
+          template <typename, typename, typename> typename Epilogue_,
+          typename TileShape, typename ClusterShape, typename KernelSchedule,
+          typename EpilogueSchedule>
+struct cutlass_3x_group_gemm {
+  using ElementAB = ElementAB_;
+  using ElementC = void;
+  using ElementD = ElementC_;
+  using ElementAccumulator = float;
+
+  using Epilogue = Epilogue_<ElementAccumulator, ElementD, TileShape>;
+
+  using StrideC =
+      cute::remove_pointer_t<cute::Stride<int64_t, cute::Int<1>, cute::Int<0>>>;
+
+  static constexpr int AlignmentAB =
+      128 / cutlass::sizeof_bits<ElementAB>::value;
+  static constexpr int AlignmentC = 128 / cutlass::sizeof_bits<ElementD>::value;
+
+  using EVTCompute = typename Epilogue::EVTCompute;
+
+  using CollectiveEpilogue =
+      typename cutlass::epilogue::collective::CollectiveBuilder<
+          ArchTag, OperatorClass, TileShape, ClusterShape,
+          cutlass::epilogue::collective::EpilogueTileAuto, ElementAccumulator,
+          ElementAccumulator, ElementC, LayoutC*, AlignmentC, ElementD,
+          LayoutC*, AlignmentC, EpilogueSchedule, EVTCompute>::CollectiveOp;
+
+  static constexpr size_t CEStorageSize =
+      sizeof(typename CollectiveEpilogue::SharedStorage);
+  using Stages = typename cutlass::gemm::collective::StageCountAutoCarveout<
+      static_cast<int>(CEStorageSize)>;
+
+  using CollectiveMainloop =
+      typename cutlass::gemm::collective::CollectiveBuilder<
+          ArchTag, OperatorClass, ElementAB, LayoutA*, AlignmentAB, ElementAB,
+          LayoutB*, AlignmentAB, ElementAccumulator, TileShape, ClusterShape,
+          Stages, KernelSchedule>::CollectiveOp;
+
+  using KernelType = enable_sm90_only<cutlass::gemm::kernel::GemmUniversal<
+      ProblemShape, CollectiveMainloop, CollectiveEpilogue>>;
+
+  struct GemmKernel : public KernelType {};
+};
+
+template <typename Gemm>
+void cutlass_group_gemm_caller(
+    torch::Tensor& out_tensors, torch::Tensor const& a_tensors,
+    torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
+    torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
+    torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
+    torch::Tensor const& b_strides, torch::Tensor const& c_strides) {
+  using ElementAB = typename Gemm::ElementAB;
+  using ElementD = typename Gemm::ElementD;
+
+  int num_experts = static_cast<int>(expert_offsets.size(0));
+  int k_size = a_tensors.size(1);
+  int n_size = out_tensors.size(1);
+
+  bool per_act_token = a_scales.numel() != 1;
+  bool per_out_ch = b_scales.numel() != num_experts;
+
+  auto stream = at::cuda::getCurrentCUDAStream(a_tensors.device().index());
+
+  auto options_int =
+      torch::TensorOptions().dtype(torch::kInt64).device(a_tensors.device());
+
+  torch::Tensor a_ptrs = torch::empty(num_experts, options_int);
+  torch::Tensor b_ptrs = torch::empty(num_experts, options_int);
+  torch::Tensor out_ptrs = torch::empty(num_experts, options_int);
+  torch::Tensor a_scales_ptrs = torch::empty(num_experts, options_int);
+  torch::Tensor b_scales_ptrs = torch::empty(num_experts, options_int);
+
+  run_get_group_gemm_starts(expert_offsets, a_ptrs, b_ptrs, out_ptrs,
+                            a_scales_ptrs, b_scales_ptrs, a_tensors, b_tensors,
+                            out_tensors, a_scales, b_scales);
+
+  using GemmKernel = typename Gemm::GemmKernel;
+  using StrideA = Stride<int64_t, Int<1>, Int<0>>;
+  using StrideB = Stride<int64_t, Int<1>, Int<0>>;
+  using StrideC = typename GemmKernel::InternalStrideC;
+
+  ProblemShape::UnderlyingProblemShape* problem_sizes_as_shapes =
+      static_cast<ProblemShape::UnderlyingProblemShape*>(
+          problem_sizes.data_ptr());
+  ProblemShape prob_shape{num_experts, problem_sizes_as_shapes, nullptr};
+
+  typename GemmKernel::MainloopArguments mainloop_args{
+      static_cast<const ElementAB**>(a_ptrs.data_ptr()),
+      static_cast<StrideA*>(a_strides.data_ptr()),
+      static_cast<const ElementAB**>(b_ptrs.data_ptr()),
+      static_cast<StrideB*>(b_strides.data_ptr())};
+
+  // Currently, we are only able to do broadcast on either all or none a_scales
+  // and on either all or none b_scales
+  typename GemmKernel::EpilogueArguments epilogue_args{
+      Gemm::Epilogue::prepare_args(
+          static_cast<const ElementAccumulator**>(a_scales_ptrs.data_ptr()),
+          static_cast<const ElementAccumulator**>(b_scales_ptrs.data_ptr()),
+          per_act_token, per_out_ch),
+      nullptr, static_cast<StrideC*>(c_strides.data_ptr()),
+      static_cast<ElementD**>(out_ptrs.data_ptr()),
+      static_cast<StrideC*>(c_strides.data_ptr())};
+
+  typename GemmKernel::Arguments args{
+      cutlass::gemm::GemmUniversalMode::kGrouped, prob_shape, mainloop_args,
+      epilogue_args};
+
+  using GemmOp = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
+  GemmOp gemm_op;
+  CUTLASS_CHECK(gemm_op.can_implement(args));
+
+  size_t workspace_size = gemm_op.get_workspace_size(args);
+  auto const workspace_options =
+      torch::TensorOptions().dtype(torch::kUInt8).device(a_tensors.device());
+  auto workspace = torch::empty(workspace_size, workspace_options);
+
+  cutlass::Status status = gemm_op.run(args, workspace.data_ptr(), stream);
+  CUTLASS_CHECK(status);
+}
+
+}  // namespace

csrc/quantization/cutlass_w8a8/moe/moe_data.cu

@@ -0,0 +1,90 @@
+#include <cudaTypedefs.h>
+
+#include <c10/cuda/CUDAGuard.h>
+#include <torch/all.h>
+
+#include <iostream>
+
+constexpr uint64_t THREADS_PER_EXPERT = 512;
+
+__global__ void compute_problem_sizes(const int* __restrict__ topk_ids,
+                                      int32_t* problem_sizes1,
+                                      int32_t* problem_sizes2,
+                                      int32_t* atomic_buffer,
+                                      const int topk_length, const int n,
+                                      const int k) {
+  int expert_id = blockIdx.x;
+
+  int occurrences = 0;
+  for (int i = threadIdx.x; i < topk_length; i += THREADS_PER_EXPERT) {
+    occurrences += (topk_ids[i] == expert_id);
+  }
+  atomicAdd(&atomic_buffer[expert_id], occurrences);
+  __syncthreads();
+
+  if (threadIdx.x == 0) {
+    int final_occurrences = atomic_buffer[expert_id];
+    problem_sizes1[expert_id * 3] = final_occurrences;
+    problem_sizes1[expert_id * 3 + 1] = 2 * n;
+    problem_sizes1[expert_id * 3 + 2] = k;
+    problem_sizes2[expert_id * 3] = final_occurrences;
+    problem_sizes2[expert_id * 3 + 1] = k;
+    problem_sizes2[expert_id * 3 + 2] = n;
+  }
+}
+
+__global__ void compute_expert_offsets(
+    const int32_t* __restrict__ problem_sizes1, int32_t* expert_offsets,
+    int32_t* atomic_buffer, const int num_experts) {
+  int32_t tot_offset = 0;
+  expert_offsets[0] = 0;
+  for (int i = 0; i < num_experts; ++i) {
+    atomic_buffer[i] = tot_offset;
+    tot_offset += problem_sizes1[i * 3];
+    expert_offsets[i + 1] = tot_offset;
+  }
+}
+
+__global__ void compute_arg_sorts(const int* __restrict__ topk_ids,
+                                  int32_t* input_permutation,
+                                  int32_t* output_permutation,
+                                  int32_t* atomic_buffer, const int topk_length,
+                                  const int topk) {
+  int expert_id = blockIdx.x;
+
+  for (int i = threadIdx.x; i < topk_length; i += THREADS_PER_EXPERT) {
+    if (topk_ids[i] == expert_id) {
+      int start = atomicAdd(&atomic_buffer[expert_id], 1);
+      input_permutation[start] = i / topk;
+      output_permutation[i] = start;
+    }
+  }
+}
+
+void get_cutlass_moe_mm_data_caller(
+    const torch::Tensor& topk_ids, torch::Tensor& expert_offsets,
+    torch::Tensor& problem_sizes1, torch::Tensor& problem_sizes2,
+    torch::Tensor& input_permutation, torch::Tensor& output_permutation,
+    const int64_t num_experts, const int64_t n, const int64_t k) {
+  auto stream = at::cuda::getCurrentCUDAStream(topk_ids.device().index());
+  auto options_int32 =
+      torch::TensorOptions().dtype(torch::kInt32).device(topk_ids.device());
+  torch::Tensor atomic_buffer = torch::zeros(num_experts, options_int32);
+
+  int num_threads = min(THREADS_PER_EXPERT, topk_ids.numel());
+  compute_problem_sizes<<<num_experts, num_threads, 0, stream>>>(
+      static_cast<const int32_t*>(topk_ids.data_ptr()),
+      static_cast<int32_t*>(problem_sizes1.data_ptr()),
+      static_cast<int32_t*>(problem_sizes2.data_ptr()),
+      static_cast<int32_t*>(atomic_buffer.data_ptr()), topk_ids.numel(), n, k);
+  compute_expert_offsets<<<1, 1, 0, stream>>>(
+      static_cast<const int32_t*>(problem_sizes1.data_ptr()),
+      static_cast<int32_t*>(expert_offsets.data_ptr()),
+      static_cast<int32_t*>(atomic_buffer.data_ptr()), num_experts);
+  compute_arg_sorts<<<num_experts, num_threads, 0, stream>>>(
+      static_cast<const int32_t*>(topk_ids.data_ptr()),
+      static_cast<int32_t*>(input_permutation.data_ptr()),
+      static_cast<int32_t*>(output_permutation.data_ptr()),
+      static_cast<int32_t*>(atomic_buffer.data_ptr()), topk_ids.numel(),
+      topk_ids.size(1));
+}

csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu

@@ -29,6 +29,20 @@ void cutlass_scaled_mm_sm90(torch::Tensor& c, torch::Tensor const& a,
                             torch::Tensor const& a_scales,
                             torch::Tensor const& b_scales,
                             std::optional<torch::Tensor> const& bias);
+
+void cutlass_moe_mm_sm90(
+    torch::Tensor& out_tensors, torch::Tensor const& a_tensors,
+    torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
+    torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
+    torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
+    torch::Tensor const& b_strides, torch::Tensor const& c_strides);
+
+void get_cutlass_moe_mm_data_caller(
+    const torch::Tensor& topk_ids, torch::Tensor& expert_offsets,
+    torch::Tensor& problem_sizes1, torch::Tensor& problem_sizes2,
+    torch::Tensor& input_permutation, torch::Tensor& output_permutation,
+    const int64_t num_experts, const int64_t n, const int64_t k);
+
 #endif
 
 #if defined ENABLE_SCALED_MM_SM100 && ENABLE_SCALED_MM_SM100
@@ -102,6 +116,19 @@ bool cutlass_scaled_mm_supports_block_fp8(int64_t cuda_device_capability) {
   return false;
 }
 
+bool cutlass_group_gemm_supported(int64_t cuda_device_capability) {
+  // CUTLASS groped FP8 kernels need at least CUDA 12.3
+  // and SM90 (Hopper)
+
+#if defined CUDA_VERSION
+  if (cuda_device_capability == 90) {
+    return CUDA_VERSION >= 12030;
+  }
+#endif
+
+  return false;
+}
+
 void cutlass_scaled_mm(torch::Tensor& c, torch::Tensor const& a,
                        torch::Tensor const& b, torch::Tensor const& a_scales,
                        torch::Tensor const& b_scales,
@@ -168,6 +195,46 @@ void cutlass_scaled_mm(torch::Tensor& c, torch::Tensor const& a,
       version_num);
 }
 
+void cutlass_moe_mm(
+    torch::Tensor& out_tensors, torch::Tensor const& a_tensors,
+    torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
+    torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
+    torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
+    torch::Tensor const& b_strides, torch::Tensor const& c_strides) {
+  int32_t version_num = get_sm_version_num();
+#if defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90
+  cutlass_moe_mm_sm90(out_tensors, a_tensors, b_tensors, a_scales, b_scales,
+                      expert_offsets, problem_sizes, a_strides, b_strides,
+                      c_strides);
+  return;
+#endif
+  TORCH_CHECK_NOT_IMPLEMENTED(
+      false,
+      "No compiled cutlass_scaled_mm for CUDA device capability: ", version_num,
+      ". Required capability: 90");
+}
+
+void get_cutlass_moe_mm_data(
+    const torch::Tensor& topk_ids, torch::Tensor& expert_offsets,
+    torch::Tensor& problem_sizes1, torch::Tensor& problem_sizes2,
+    torch::Tensor& input_permutation, torch::Tensor& output_permutation,
+    const int64_t num_experts, const int64_t n, const int64_t k) {
+  // This function currently gets compiled only if we have a valid cutlass moe
+  // mm to run it for.
+  int32_t version_num = get_sm_version_num();
+#if defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90
+  get_cutlass_moe_mm_data_caller(topk_ids, expert_offsets, problem_sizes1,
+                                 problem_sizes2, input_permutation,
+                                 output_permutation, num_experts, n, k);
+  return;
+#endif
+  TORCH_CHECK_NOT_IMPLEMENTED(
+      false,
+      "No compiled get_cutlass_moe_mm_data: no cutlass_scaled_mm kernel for "
+      "CUDA device capability: ",
+      version_num, ". Required capability: 90");
+}
+
 void cutlass_scaled_mm_azp(torch::Tensor& c, torch::Tensor const& a,
                            torch::Tensor const& b,
                            torch::Tensor const& a_scales,

csrc/quantization/fp8/common.cu

@@ -30,9 +30,6 @@ __global__ void dynamic_per_token_scaled_fp8_quant_kernel(
     fp8_type* __restrict__ out, float* __restrict__ scale,
     scalar_t const* __restrict__ input, float const* __restrict__ scale_ub,
     const int hidden_size) {
-  float const min_scaling_factor =
-      1.0f / (fp8_e4m3_adjusted_max_v<fp8_type> * 512.f);
-
   int const tid = threadIdx.x;
   int const token_idx = blockIdx.x;
 
@@ -67,8 +64,8 @@ __global__ void dynamic_per_token_scaled_fp8_quant_kernel(
       token_scale = block_absmax_val_maybe;
     }
     // token scale computation
-    token_scale = max(token_scale / fp8_e4m3_adjusted_max_v<fp8_type>,
-                      min_scaling_factor);
+    token_scale = max(token_scale / quant_type_max_v<fp8_type>,
+                      min_scaling_factor<fp8_type>::val());
     scale[token_idx] = token_scale;
   }
   __syncthreads();

csrc/quantization/fp8/common.cuh

@@ -1,20 +1,12 @@
 #pragma once
 
 #include "quantization/vectorization.cuh"
+#include "quantization/utils.cuh"
 
 #include <cmath>
-#include <c10/core/ScalarType.h>
 
-#ifndef USE_ROCM
-  #include <c10/util/Float8_e4m3fn.h>
-  #define MAYBE_HOST_DEVICE C10_HOST_DEVICE
-#else
-  #include <ATen/hip/HIPContext.h>
-  #include <c10/util/Float8_e4m3fn.h>
-  #include <c10/util/Float8_e4m3fnuz.h>
+#ifdef USE_ROCM
   #include "amd/quant_utils.cuh"
-  // ROCm doesn't seem to need C10_HOST_DEVICE for static constexpr
-  #define MAYBE_HOST_DEVICE
 #endif
 
 // Determines the preferred FP8 type for the current platform.
@@ -31,29 +23,6 @@ static bool is_fp8_ocp() {
 #endif
 }
 
-template <typename T>
-struct fp8_e4m3_adjusted_max;
-
-template <>
-struct fp8_e4m3_adjusted_max<c10::Float8_e4m3fn> {
-  static constexpr c10::Float8_e4m3fn val() {
-    return std::numeric_limits<c10::Float8_e4m3fn>::max();
-  }
-};
-
-// Using the default max value from pytorch (240.0 0x7F) will cause accuracy
-// issues when running dynamic quantization. Here use 224.0 0x7E for rocm.
-template <>
-struct fp8_e4m3_adjusted_max<c10::Float8_e4m3fnuz> {
-  static constexpr c10::Float8_e4m3fnuz val() {
-    return c10::Float8_e4m3fnuz(0x7E, c10::Float8_e4m3fnuz::from_bits());
-  }
-};
-
-template <typename T>
-MAYBE_HOST_DEVICE static constexpr T fp8_e4m3_adjusted_max_v =
-    fp8_e4m3_adjusted_max<T>::val();
-
 namespace vllm {
 
 __device__ __forceinline__ float atomicMaxFloat(float* addr, float value) {
@@ -76,8 +45,8 @@ __device__ __forceinline__ fp8_type scaled_fp8_conversion(float const val,
     x = val / scale;
   }
 
-  float r = fmax(-fp8_e4m3_adjusted_max_v<fp8_type>,
-                 fmin(x, fp8_e4m3_adjusted_max_v<fp8_type>));
+  float r =
+      fmax(-quant_type_max_v<fp8_type>, fmin(x, quant_type_max_v<fp8_type>));
 #ifndef USE_ROCM
   return static_cast<fp8_type>(r);
 #else
@@ -123,7 +92,7 @@ __global__ void segmented_max_reduction(float* __restrict__ scale,
   // Finally, since cache[0] contains the maximum for this thread block,
   // atomically write the max to the target location
   if (threadIdx.x == 0) {
-    atomicMaxFloat(scale, cache[0] / fp8_e4m3_adjusted_max_v<fp8_type>);
+    atomicMaxFloat(scale, cache[0] / quant_type_max_v<fp8_type>);
   }
 }
 

csrc/quantization/fused_kernels/fused_layernorm_dynamic_per_token_quant.cu

@@ -14,8 +14,7 @@ __device__ void rms_norm_dynamic_per_token_quant_vec(
     float* __restrict__ scales,           // [num_tokens]
     scalar_t const* __restrict__ input,   // [..., hidden_size]
     scalar_t const* __restrict__ weight,  // [hidden_size]
-    float const* scale_ub, float const var_epsilon,
-    float const min_scaling_factor, int32_t const hidden_size,
+    float const* scale_ub, float const var_epsilon, int32_t const hidden_size,
     scalar_t* __restrict__ residual = nullptr) {
   float rms = 0.0f;
   float token_scale = 0.0f;
@@ -27,8 +26,8 @@ __device__ void rms_norm_dynamic_per_token_quant_vec(
   // Compute scale
   vllm::vectorized::compute_dynamic_per_token_scales<scalar_t, scalar_out_t,
                                                      has_residual>(
-      &token_scale, scales, input, weight, rms, scale_ub, min_scaling_factor,
-      hidden_size, residual);
+      &token_scale, scales, input, weight, rms, scale_ub, hidden_size,
+      residual);
 
   // RMS Norm + Quant
   if constexpr (std::is_same_v<scalar_out_t, int8_t>) {
@@ -50,8 +49,7 @@ __global__ void rms_norm_dynamic_per_token_quant_kernel(
     float* __restrict__ scales,           // [num_tokens]
     scalar_t const* __restrict__ input,   // [..., hidden_size]
     scalar_t const* __restrict__ weight,  // [hidden_size]
-    float const* scale_ub, float const var_epsilon,
-    float const min_scaling_factor, int32_t const hidden_size,
+    float const* scale_ub, float const var_epsilon, int32_t const hidden_size,
     scalar_t* __restrict__ residual = nullptr) {
   // For vectorization, token_input and token_output pointers need to be
   // aligned at 8-byte and 4-byte addresses respectively.
@@ -60,8 +58,8 @@ __global__ void rms_norm_dynamic_per_token_quant_kernel(
   if (can_vectorize) {
     return rms_norm_dynamic_per_token_quant_vec<scalar_t, scalar_out_t,
                                                 has_residual>(
-        out, scales, input, weight, scale_ub, var_epsilon, min_scaling_factor,
-        hidden_size, residual);
+        out, scales, input, weight, scale_ub, var_epsilon, hidden_size,
+        residual);
   }
 
   float rms = 0.0f;
@@ -72,8 +70,8 @@ __global__ void rms_norm_dynamic_per_token_quant_kernel(
                                             var_epsilon, residual);
   // Compute Scale
   vllm::compute_dynamic_per_token_scales<scalar_t, scalar_out_t, has_residual>(
-      &token_scale, scales, input, weight, rms, scale_ub, min_scaling_factor,
-      hidden_size, residual);
+      &token_scale, scales, input, weight, rms, scale_ub, hidden_size,
+      residual);
 
   // RMS Norm + Quant
   if constexpr (std::is_same_v<scalar_out_t, int8_t>) {
@@ -105,11 +103,6 @@ void rms_norm_dynamic_per_token_quant_dispatch(
   const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
 
-  const float min_scaling_factor =
-      out.dtype() == torch::kInt8
-          ? std::numeric_limits<float>::epsilon()
-          : 1.0f / (std::numeric_limits<c10::Float8_e4m3fn>::max() * 512.f);
-
   if (residual.has_value()) {
     VLLM_DISPATCH_QUANT_TYPES(
         out.scalar_type(), "rms_norm_dynamic_per_token_quant_kernel", [&] {
@@ -119,8 +112,7 @@ void rms_norm_dynamic_per_token_quant_dispatch(
                   out.data_ptr<scalar_t>(), scales.data_ptr<float>(),
                   input.data_ptr<scalar_in_t>(), weight.data_ptr<scalar_in_t>(),
                   scale_ub.has_value() ? scale_ub->data_ptr<float>() : nullptr,
-                  var_epsilon, min_scaling_factor, hidden_size,
-                  residual->data_ptr<scalar_in_t>());
+                  var_epsilon, hidden_size, residual->data_ptr<scalar_in_t>());
         });
 
   } else {
@@ -132,7 +124,7 @@ void rms_norm_dynamic_per_token_quant_dispatch(
                   out.data_ptr<scalar_t>(), scales.data_ptr<float>(),
                   input.data_ptr<scalar_in_t>(), weight.data_ptr<scalar_in_t>(),
                   scale_ub.has_value() ? scale_ub->data_ptr<float>() : nullptr,
-                  var_epsilon, min_scaling_factor, hidden_size, nullptr);
+                  var_epsilon, hidden_size, nullptr);
         });
   }
 }

csrc/quantization/fused_kernels/layernorm_utils.cuh

@@ -5,6 +5,7 @@
  */
 
 #include "quantization/vectorization.cuh"
+#include "quantization/utils.cuh"
 #include "quant_conversions.cuh"
 
 #ifndef USE_ROCM
@@ -24,7 +25,7 @@ __device__ void compute_rms(float* rms, scalar_t const* __restrict__ input,
   // sum of squares
   float ss = 0.0f;
 
-  for (int32_t i = threadIdx.x; i < hidden_size; i += blockDim.x) {
+  for (auto i = threadIdx.x; i < hidden_size; i += blockDim.x) {
     float x = static_cast<float>(input[token_offset + i]);
     if constexpr (has_residual) {
       x += static_cast<float>(residual[token_offset + i]);
@@ -51,14 +52,14 @@ __device__ void compute_dynamic_per_token_scales(
     float* __restrict__ token_scale, float* __restrict__ all_token_scales,
     scalar_t const* __restrict__ input, scalar_t const* __restrict__ weight,
     float const rms, float const* __restrict__ scale_ub,
-    float const min_scaling_factor, int32_t const hidden_size,
+    int32_t const hidden_size,
     scalar_t const* __restrict__ residual = nullptr) {
   int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
   ;
-  constexpr scalar_out_t qmax{std::numeric_limits<scalar_out_t>::max()};
+  constexpr scalar_out_t qmax{quant_type_max_v<scalar_out_t>};
 
   float block_absmax_val_maybe = 0.0f;
-  for (int32_t i = threadIdx.x; i < hidden_size; i += blockDim.x) {
+  for (auto i = threadIdx.x; i < hidden_size; i += blockDim.x) {
     float x = static_cast<float>(input[token_offset + i]);
     if constexpr (has_residual) {
       x += static_cast<float>(residual[token_offset + i]);
@@ -83,7 +84,7 @@ __device__ void compute_dynamic_per_token_scales(
       scale = block_absmax_val_maybe;
     }
     // token scale computation
-    scale = max(scale / qmax, min_scaling_factor);
+    scale = max(scale / qmax, min_scaling_factor<scalar_out_t>::val());
     s_token_scale = scale;                 // Shared memory store
     all_token_scales[blockIdx.x] = scale;  // Global output store
   }
@@ -103,7 +104,7 @@ __device__ void norm_and_quant(scalar_out_t* __restrict__ output,
   int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
   ;
 
-  for (int32_t i = threadIdx.x; i < hidden_size; i += blockDim.x) {
+  for (auto i = threadIdx.x; i < hidden_size; i += blockDim.x) {
     float x = static_cast<float>(input[token_offset + i]);
     if constexpr (has_residual) {
       x += static_cast<float>(residual[token_offset + i]);
@@ -142,7 +143,7 @@ __device__ void compute_rms(float* rms, scalar_t const* __restrict__ input,
   int32_t const num_vec_elems = hidden_size >> 2;
 
 #pragma unroll 4
-  for (int32_t i = threadIdx.x; i < num_vec_elems; i += blockDim.x) {
+  for (auto i = threadIdx.x; i < num_vec_elems; i += blockDim.x) {
     vec4_t<scalar_t> in = vec_input[i];
 
     vec4_t<float> x;
@@ -184,7 +185,7 @@ __device__ void compute_dynamic_per_token_scales(
     float* __restrict__ token_scale, float* __restrict__ all_token_scales,
     scalar_t const* __restrict__ input, scalar_t const* __restrict__ weight,
     float const rms, float const* __restrict__ scale_ub,
-    float const min_scaling_factor, int32_t const hidden_size,
+    int32_t const hidden_size,
     scalar_t const* __restrict__ residual = nullptr) {
   int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
   ;
@@ -200,13 +201,13 @@ __device__ void compute_dynamic_per_token_scales(
         reinterpret_cast<vec4_t<scalar_t> const*>(&residual[token_offset]);
   }
 
-  constexpr scalar_out_t qmax{std::numeric_limits<scalar_out_t>::max()};
+  constexpr scalar_out_t qmax{quant_type_max_v<scalar_out_t>};
 
   int32_t const num_vec_elems = hidden_size >> 2;
   float block_absmax_val_maybe = 0.0f;
 
 #pragma unroll 4
-  for (int32_t i = threadIdx.x; i < num_vec_elems; i += blockDim.x) {
+  for (auto i = threadIdx.x; i < num_vec_elems; i += blockDim.x) {
     vec4_t<scalar_t> in = vec_input[i];
     vec4_t<scalar_t> const w = vec_weight[i];
 
@@ -248,7 +249,7 @@ __device__ void compute_dynamic_per_token_scales(
       scale = block_absmax_val_maybe;
     }
     // token scale computation
-    scale = max(scale / qmax, min_scaling_factor);
+    scale = max(scale / qmax, min_scaling_factor<scalar_out_t>::val());
     s_token_scale = scale;                 // shared memory store
     all_token_scales[blockIdx.x] = scale;  // global output store
   }
@@ -286,7 +287,7 @@ __device__ void norm_and_quant(scalar_out_t* __restrict__ output,
 // TODO(luka/varun) extract into type-agnostic vectorized quant function to
 //  replace scaled_fp8_conversion_vec
 #pragma unroll 4
-  for (int32_t i = threadIdx.x; i < num_vec_elems; i += blockDim.x) {
+  for (auto i = threadIdx.x; i < num_vec_elems; i += blockDim.x) {
     vec4_t<scalar_t> const in = vec_input[i];
     vec4_t<scalar_t> const w = vec_weight[i];
 

csrc/quantization/fused_kernels/quant_conversions.cuh

@@ -33,8 +33,8 @@ static __device__ __forceinline__ int8_t float_to_int8_rn(float const x) {
 
 template <typename fp8_type>
 static __device__ __forceinline__ fp8_type float_to_fp8(float const x) {
-  float const r = fmax(-fp8_e4m3_adjusted_max_v<fp8_type>,
-                       fmin(x, fp8_e4m3_adjusted_max_v<fp8_type>));
+  float const r =
+      fmax(-quant_type_max_v<fp8_type>, fmin(x, quant_type_max_v<fp8_type>));
   return static_cast<fp8_type>(r);
 }
 

csrc/quantization/gguf/dequantize.cuh

@@ -94,17 +94,17 @@ static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __
     dfloat2 v;
     dequantize_kernel(vx, ib, iqs, v);
 
-    y[iybs + iqs + 0]        = v.x;
-    y[iybs + iqs + y_offset] = v.y;
+    y[iybs + iqs + 0]        = convert_from_half<dst_t>(v.x);
+    y[iybs + iqs + y_offset] = convert_from_half<dst_t>(v.y);
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
-    const int i   = blockIdx.x;
+    const auto i   = blockIdx.x;
     const block_q2_K * x = (const block_q2_K *) vx;
 
-    const int tid = threadIdx.x;
+    const auto tid = threadIdx.x;
     const int n   = tid/32;
     const int l   = tid - 32*n;
     const int is  = 8*n + l/16;
@@ -114,19 +114,19 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t
 
     half dall = __low2half(x[i].dm);
     half dmin = __high2half(x[i].dm);
-    y[l+ 0] = __hsub(__hmul(dall, __int2half_rn((x[i].scales[is+0] & 0xF) * ((q >> 0) & 3))), __hmul(dmin,  __int2half_rn(x[i].scales[is+0] >> 4)));
-    y[l+32] = __hsub(__hmul(dall, __int2half_rn((x[i].scales[is+2] & 0xF) * ((q >> 2) & 3))), __hmul(dmin,  __int2half_rn(x[i].scales[is+2] >> 4)));
-    y[l+64] = __hsub(__hmul(dall, __int2half_rn((x[i].scales[is+4] & 0xF) * ((q >> 4) & 3))), __hmul(dmin,  __int2half_rn(x[i].scales[is+4] >> 4)));
-    y[l+96] = __hsub(__hmul(dall, __int2half_rn((x[i].scales[is+6] & 0xF) * ((q >> 6) & 3))), __hmul(dmin,  __int2half_rn(x[i].scales[is+6] >> 4)));
+    y[l+ 0] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+0] & 0xF) * ((q >> 0) & 3))), __hmul(dmin,  __int2half_rn(x[i].scales[is+0] >> 4))));
+    y[l+32] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+2] & 0xF) * ((q >> 2) & 3))), __hmul(dmin,  __int2half_rn(x[i].scales[is+2] >> 4))));
+    y[l+64] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+4] & 0xF) * ((q >> 4) & 3))), __hmul(dmin,  __int2half_rn(x[i].scales[is+4] >> 4))));
+    y[l+96] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+6] & 0xF) * ((q >> 6) & 3))), __hmul(dmin,  __int2half_rn(x[i].scales[is+6] >> 4))));
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
-    const int i = blockIdx.x;
+    const auto i = blockIdx.x;
     const block_q3_K * x = (const block_q3_K *) vx;
 
-    const int r = threadIdx.x/4;
+    const auto r = threadIdx.x/4;
     const int tid = r/2;
     const int is0 = r%2;
     const int l0 = 16*is0 + 4*(threadIdx.x%4);
@@ -148,7 +148,9 @@ static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t
     const uint8_t * q = x[i].qs + 32*n;
     const uint8_t * hm = x[i].hmask;
 
-    for (int l = l0; l < l0+4; ++l) y[l] = __hmul(dl,  __int2half_rn((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4)));
+    for (int l = l0; l < l0+4; ++l) {
+        y[l] = convert_from_half<dst_t>(__hmul(dl,  __int2half_rn((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4))));
+    }
 }
 
 static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
@@ -164,10 +166,10 @@ template<typename dst_t>
 static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
     const block_q4_K * x = (const block_q4_K *) vx;
 
-    const int i = blockIdx.x;
+    const auto i = blockIdx.x;
 
     // assume 32 threads
-    const int tid = threadIdx.x;
+    const auto tid = threadIdx.x;
     const int il  = tid/8;
     const int ir  = tid%8;
     const int is  = 2*il;
@@ -188,8 +190,8 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t
     const half d2 = __hmul(dall, __int2half_rn(sc));
     const half m2 = __hmul(dmin, __int2half_rn(m));
     for (int l = 0; l < n; ++l) {
-        y[l + 0] = __hsub(__hmul(d1, __int2half_rn(q[l] & 0xF)), m1);
-        y[l +32] = __hsub(__hmul(d2,  __int2half_rn(q[l] >> 4)), m2);
+        y[l + 0] = convert_from_half<dst_t>(__hsub(__hmul(d1, __int2half_rn(q[l] & 0xF)), m1));
+        y[l +32] = convert_from_half<dst_t>(__hsub(__hmul(d2,  __int2half_rn(q[l] >> 4)), m2));
     }
 }
 
@@ -197,10 +199,10 @@ template<typename dst_t>
 static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
     const block_q5_K * x = (const block_q5_K *) vx;
 
-    const int i = blockIdx.x;
+    const auto i = blockIdx.x;
 
     // assume 64 threads - this is very slightly better than the one below
-    const int tid = threadIdx.x;
+    const auto tid = threadIdx.x;
     const int il  = tid/16;   // il is in 0...3
     const int ir  = tid%16;   // ir is in 0...15
     const int is  = 2*il;     // is is in 0...6
@@ -220,21 +222,21 @@ static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t
     const half d2 = __hmul(dall, __int2half_rn(sc)); const half m2 = __hmul(dmin, __int2half_rn(m));
 
     uint8_t   hm  = 1 << (2*il);
-    y[ 0] = __hsub(__hmul(d1, __int2half_rn((ql[0] & 0xF) + (qh[0] & hm ? 16 : 0))), m1);
-    y[ 1] = __hsub(__hmul(d1, __int2half_rn((ql[1] & 0xF) + (qh[1] & hm ? 16 : 0))), m1);
+    y[ 0] = convert_from_half<dst_t>(__hsub(__hmul(d1, __int2half_rn((ql[0] & 0xF) + (qh[0] & hm ? 16 : 0))), m1));
+    y[ 1] = convert_from_half<dst_t>(__hsub(__hmul(d1, __int2half_rn((ql[1] & 0xF) + (qh[1] & hm ? 16 : 0))), m1));
     hm <<= 1;
-    y[32] = __hsub(__hmul(d2, __int2half_rn((ql[0] >>  4) + (qh[0] & hm ? 16 : 0))), m2);
-    y[33] = __hsub(__hmul(d2, __int2half_rn((ql[1] >>  4) + (qh[1] & hm ? 16 : 0))), m2);
+    y[32] = convert_from_half<dst_t>(__hsub(__hmul(d2, __int2half_rn((ql[0] >>  4) + (qh[0] & hm ? 16 : 0))), m2));
+    y[33] = convert_from_half<dst_t>(__hsub(__hmul(d2, __int2half_rn((ql[1] >>  4) + (qh[1] & hm ? 16 : 0))), m2));
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
     const block_q6_K * x = (const block_q6_K *) vx;
 
-    const int i = blockIdx.x;
+    const auto i = blockIdx.x;
 
     // assume 64 threads - this is very slightly better than the one below
-    const int tid = threadIdx.x;
+    const auto tid = threadIdx.x;
     const int ip  = tid/32;   // ip is 0 or 1
     const int il  = tid - 32*ip; // 0...32
     const int is  = 8*ip + il/16;
@@ -247,19 +249,19 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t
     const uint8_t   qh = x[i].qh[32*ip + il];
     const int8_t  * sc = x[i].scales + is;
 
-    y[ 0] = __hmul(d, __int2half_rn(sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32)));
-    y[32] = __hmul(d, __int2half_rn(sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32)));
-    y[64] = __hmul(d, __int2half_rn(sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh >> 4) & 3) << 4)) - 32)));
-    y[96] = __hmul(d, __int2half_rn(sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32)));
+    y[ 0] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32))));
+    y[32] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32))));
+    y[64] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh >> 4) & 3) << 4)) - 32))));
+    y[96] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32))));
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
-    const int i   = blockIdx.x;
+    const auto i   = blockIdx.x;
     const block_iq2_xxs * x = (const block_iq2_xxs  *) vx;
 
-    const int tid = threadIdx.x;
+    const auto tid = threadIdx.x;
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -269,16 +271,16 @@ static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, ds
     const uint32_t aux32 = q2[2] | (q2[3] << 16);
     const float d = __half2float(x[i].d) * (0.5f + (aux32 >> 28)) * 0.25f;
     const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
-    for (int j = 0; j < 8; ++j) y[j] = __float2half(d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f));
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
-    const int i   = blockIdx.x;
+    const auto i   = blockIdx.x;
     const block_iq2_xs * x = (const block_iq2_xs *) vx;
 
-    const int tid = threadIdx.x;
+    const auto tid = threadIdx.x;
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -286,33 +288,33 @@ static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst
     const uint8_t  * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511));
     const float d = __half2float(x[i].d) * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
     const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
-    for (int j = 0; j < 8; ++j) y[j] = __float2half(d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f));
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
 
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
-    const int i   = blockIdx.x;
+    const auto i   = blockIdx.x;
     const block_iq2_s * x = (const block_iq2_s *) vx;
 
-    const int tid = threadIdx.x;
+    const auto tid = threadIdx.x;
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
     const uint8_t * grid = (const uint8_t *)(iq2s_grid + (x[i].qs[4*ib+il] | ((x[i].qh[ib] << (8-2*il)) & 0x300)));
     const float d = __half2float(x[i].d) * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
     const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
-    for (int j = 0; j < 8; ++j) y[j] = __float2half(d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f));
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
-    const int i   = blockIdx.x;
+    const auto i   = blockIdx.x;
     const block_iq3_xxs * x = (const block_iq3_xxs  *) vx;
 
-    const int tid = threadIdx.x;
+    const auto tid = threadIdx.x;
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -324,18 +326,18 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds
     const float d = __half2float(x[i].d) * (0.5f + (aux32 >> 28)) * 0.5f;
     const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
     for (int j = 0; j < 4; ++j) {
-        y[j+0] = __float2half(d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f));
-        y[j+4] = __float2half(d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f));
+        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
+        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
     }
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
-    const int i   = blockIdx.x;
+    const auto i   = blockIdx.x;
     const block_iq3_s * x = (const block_iq3_s *) vx;
 
-    const int tid = threadIdx.x;
+    const auto tid = threadIdx.x;
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -345,8 +347,8 @@ static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_
     const float d = __half2float(x[i].d) * (0.5f + ((x[i].scales[ib/2] >> 4*(ib%2)) & 0xf)) * 0.5f;
     const uint8_t signs = x[i].signs[4*ib + il];
     for (int j = 0; j < 4; ++j) {
-        y[j+0] = __float2half(d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f));
-        y[j+4] = __float2half(d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f));
+        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
+        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
     }
 }
 
@@ -367,7 +369,7 @@ static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_
     grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
     grid32[0] &= 0x0f0f0f0f;
     for (int j = 0; j < 8; ++j) {
-        y[j] = __float2half(d * (q[j] + delta));
+        y[j] = d * (q[j] + delta);
     }
 }
 
@@ -392,43 +394,43 @@ static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_
     grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
     grid32[0] &= 0x0f0f0f0f;
     for (int j = 0; j < 8; ++j) {
-        y[j] = __float2half(d * (q[j] + delta));
+        y[j] = d * (q[j] + delta);
     }
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
-    const int i   = blockIdx.x;
+    const auto i   = blockIdx.x;
     const block_iq4_nl * x = (const block_iq4_nl *) vx + i*(QK_K/QK4_NL);
 
-    const int tid = threadIdx.x;
+    const auto tid = threadIdx.x;
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 4*il;
     const uint8_t  * q4 = x[ib].qs + 4*il;
     const float d = __half2float(x[ib].d);
     for (int j = 0; j < 4; ++j) {
-        y[j+ 0] = __float2half(d * kvalues_iq4nl[q4[j] & 0xf]);
-        y[j+16] = __float2half(d * kvalues_iq4nl[q4[j] >>  4]);
+        y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
+        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
     }
 
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-    const int i   = blockIdx.x;
+    const auto i   = blockIdx.x;
     const block_iq4_xs * x = (const block_iq4_xs *)vx;
 
-    const int tid = threadIdx.x;
+    const auto tid = threadIdx.x;
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 4*il;
     const uint8_t  * q4 = x[i].qs + 16*ib + 4*il;
     const float d = __half2float(x[i].d) * ((((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4)) - 32);
     for (int j = 0; j < 4; ++j) {
-        y[j+ 0] = __float2half(d * kvalues_iq4nl[q4[j] & 0xf]);
-        y[j+16] = __float2half(d * kvalues_iq4nl[q4[j] >>  4]);
+        y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
+        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
     }
 }
 
@@ -522,7 +524,8 @@ static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int k,
     dequantize_block_iq4_xs<<<nb, 32, 0, stream>>>(vx, y);
 }
 
-static to_fp16_cuda_t ggml_get_to_fp16_cuda(int64_t type) {
+template<typename dst_t>
+static to_cuda_ggml_t<dst_t> ggml_get_to_cuda(int64_t type) {
     switch (type) {
         case 2:
             return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
@@ -565,4 +568,4 @@ static to_fp16_cuda_t ggml_get_to_fp16_cuda(int64_t type) {
         default:
             return nullptr;
     }
-}
+}

csrc/quantization/gguf/ggml-common.h

@@ -1063,7 +1063,8 @@ static const __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -
 typedef half dfloat; // dequantize float
 typedef half2 dfloat2;
 typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, dfloat2 & v);
-typedef void (*to_fp16_cuda_t)(const void * __restrict__ x, dfloat * __restrict__ y, int k, cudaStream_t stream);
+template<typename dst_t>
+using to_cuda_ggml_t = void (*)(const void * __restrict__ x, dst_t * __restrict__ y, int k, cudaStream_t stream);
 typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs);
 typedef void (*allocate_tiles_cuda_t)(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc);
 typedef void (*load_tiles_cuda_t)(
@@ -1075,6 +1076,25 @@ typedef float (*vec_dot_q_mul_mat_cuda_t)(
 
 // Utility function
 
+template<typename dst_t>
+static __device__ __forceinline__ dst_t convert_from_half(half val) {
+    return val;
+}
+
+template<>
+__device__ __forceinline__ c10::BFloat16 convert_from_half<c10::BFloat16>(half val) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
+    return __float2bfloat16(__half2float(val));
+#else
+    return __half2float(val);
+#endif  // defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
+}
+
+template<>
+__device__ __forceinline__ float convert_from_half<float>(half val) {
+    return __half2float(val);
+}
+
 #if defined(USE_ROCM)
 
 #ifndef __has_builtin

csrc/quantization/gguf/gguf_kernel.cu

@@ -19,11 +19,11 @@ template <typename scalar_t>
 static __global__ void quantize_q8_1(const scalar_t* __restrict__ x,
                                      void* __restrict__ vy, const int kx,
                                      const int kx_padded) {
-  const int ix = blockDim.x * blockIdx.x + threadIdx.x;
+  const auto ix = blockDim.x * blockIdx.x + threadIdx.x;
   if (ix >= kx_padded) {
     return;
   }
-  const int iy = blockDim.y * blockIdx.y + threadIdx.y;
+  const auto iy = blockDim.y * blockIdx.y + threadIdx.y;
   const int i_padded = iy * kx_padded + ix;
 
   block_q8_1* y = (block_q8_1*)vy;
@@ -71,14 +71,19 @@ static void quantize_row_q8_1_cuda(const scalar_t* x, void* vy, const int kx,
 }
 
 torch::Tensor ggml_dequantize(torch::Tensor W,  // quant weight
-                              int64_t type, int64_t m, int64_t n) {
+                              int64_t type, int64_t m, int64_t n,
+                              std::optional<at::ScalarType> const& dtype) {
   const at::cuda::OptionalCUDAGuard device_guard(device_of(W));
-  auto options =
-      torch::TensorOptions().dtype(torch::kFloat16).device(W.device());
+  auto dtype_ = dtype.value_or(torch::kFloat16);
+  auto options = torch::TensorOptions().dtype(dtype_).device(W.device());
   at::Tensor DW = torch::empty({m, n}, options);
   cudaStream_t stream = at::cuda::getCurrentCUDAStream().stream();
-  const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(type);
-  to_fp16_cuda((void*)W.data_ptr(), (half*)DW.data_ptr(), m * n, stream);
+
+  VLLM_DISPATCH_FLOATING_TYPES(DW.scalar_type(), "ggml_dequantize", [&] {
+    auto to_cuda = ggml_get_to_cuda<scalar_t>(type);
+    to_cuda((void*)W.data_ptr(), (scalar_t*)DW.data_ptr(), m * n, stream);
+  });
+
   return DW;
 }
 
@@ -375,25 +380,25 @@ torch::Tensor ggml_moe_a8(torch::Tensor X,  // input
 int64_t ggml_moe_get_block_size(int64_t type) {
   switch (type) {
     case 2:
-      return MMQ_X_Q4_0;
+      return MOE_X_Q4_0;
     case 3:
-      return MMQ_X_Q4_1;
+      return MOE_X_Q4_1;
     case 6:
-      return MMQ_X_Q5_0;
+      return MOE_X_Q5_0;
     case 7:
-      return MMQ_X_Q5_1;
+      return MOE_X_Q5_1;
     case 8:
-      return MMQ_X_Q8_0;
+      return MOE_X_Q8_0;
     case 10:
-      return MMQ_X_Q2_K;
+      return MOE_X_Q2_K;
     case 11:
-      return MMQ_X_Q3_K;
+      return MOE_X_Q3_K;
     case 12:
-      return MMQ_X_Q4_K;
+      return MOE_X_Q4_K;
     case 13:
-      return MMQ_X_Q5_K;
+      return MOE_X_Q5_K;
     case 14:
-      return MMQ_X_Q6_K;
+      return MOE_X_Q6_K;
   }
   return 0;
 }

csrc/quantization/gguf/mmq.cuh

@@ -14,10 +14,10 @@ static __device__ __forceinline__ void mul_mat_q(
 
     const int & ncols_dst = ncols_y;
 
-    const int row_dst_0 = blockIdx.x*mmq_y;
+    const auto row_dst_0 = blockIdx.x*mmq_y;
     const int & row_x_0 = row_dst_0;
 
-    const int col_dst_0 = blockIdx.y*mmq_x;
+    const auto col_dst_0 = blockIdx.y*mmq_x;
     const int & col_y_0 = col_dst_0;
 
     int   * tile_x_ql = nullptr;
@@ -39,7 +39,7 @@ static __device__ __forceinline__ void mul_mat_q(
 
 #pragma unroll
         for (int ir = 0; ir < qr && ib0 + ir * blocks_per_warp/qr < blocks_per_row_x; ++ir) {
-            const int kqs = ir*WARP_SIZE_GGUF + threadIdx.x;
+            const auto kqs = ir*WARP_SIZE_GGUF + threadIdx.x;
             const int kbxd = kqs / QI8_1;
 
 #pragma unroll
@@ -53,7 +53,7 @@ static __device__ __forceinline__ void mul_mat_q(
 #pragma unroll
             for (int ids0 = 0; ids0 < mmq_x; ids0 += nwarps * QI8_1) {
                 const int ids = (ids0 + threadIdx.y * QI8_1 + threadIdx.x / (WARP_SIZE_GGUF/QI8_1)) % mmq_x;
-                const int kby = threadIdx.x % (WARP_SIZE_GGUF/QI8_1);
+                const auto kby = threadIdx.x % (WARP_SIZE_GGUF/QI8_1);
                 const int col_y_eff = min(col_y_0 + ids, ncols_y-1);
 
                 // if the sum is not needed it's faster to transform the scale to f32 ahead of time
@@ -87,14 +87,14 @@ static __device__ __forceinline__ void mul_mat_q(
 
 #pragma unroll
     for (int j = 0; j < mmq_x; j += nwarps) {
-        const int col_dst = col_dst_0 + j + threadIdx.y;
+        const auto col_dst = col_dst_0 + j + threadIdx.y;
         if (col_dst >= ncols_dst) {
             return;
         }
 
 #pragma unroll
         for (int i = 0; i < mmq_y; i += WARP_SIZE_GGUF) {
-            const int row_dst = row_dst_0 + threadIdx.x + i;
+            const auto row_dst = row_dst_0 + threadIdx.x + i;
             if (row_dst >= nrows_dst) {
                 continue;
             }

csrc/quantization/gguf/mmvq.cuh

@@ -1,7 +1,7 @@
 // copied and adapted from https://github.com/ggerganov/llama.cpp/blob/b2899/ggml-cuda/mmvq.cu
 template <typename scalar_t, int qk, int qi, typename block_q_t, int vdr, vec_dot_q_cuda_t vec_dot_q_cuda>
 static __global__ void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict__ vy, scalar_t * __restrict__ dst, const int ncols, const int nrows) {
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
+    const auto row = blockIdx.x*blockDim.y + threadIdx.y;
 
     if (row >= nrows) {
         return;
@@ -16,7 +16,7 @@ static __global__ void mul_mat_vec_q(const void * __restrict__ vx, const void *
     const block_q_t  * x = (const block_q_t  *) vx;
     const block_q8_1 * y = (const block_q8_1 *) vy;
 
-    for (int i = threadIdx.x / (qi/vdr); i < blocks_per_row; i += blocks_per_warp) {
+    for (auto i = threadIdx.x / (qi/vdr); i < blocks_per_row; i += blocks_per_warp) {
         const int ibx = row*blocks_per_row + i; // x block index
 
         const int iby = i * (qk/QK8_1); // y block index that aligns with ibx

csrc/quantization/gguf/moe.cuh

@@ -19,10 +19,10 @@ static __device__ __forceinline__ void moe_q(
 
   const int ncols_dst = ncols_y * top_k;
 
-  const int row_dst_0 = blockIdx.x * mmq_y;
+  const auto row_dst_0 = blockIdx.x * mmq_y;
   const int& row_x_0 = row_dst_0;
 
-  const int col_dst_0 = blockIdx.y * mmq_x;
+  const auto col_dst_0 = blockIdx.y * mmq_x;
 
   int token_offs[mmq_x / nwarps];
   for (int i = 0; i < mmq_x; i += nwarps) {
@@ -56,7 +56,7 @@ static __device__ __forceinline__ void moe_q(
     const int n_per_r = ((qk * blocks_per_warp) / qr);
 #pragma unroll
     for (int ir = 0; ir < qr && ib0 * qk + ir * n_per_r < ncols_x; ++ir) {
-      const int kqs = ir * WARP_SIZE_GGUF + threadIdx.x;
+      const auto kqs = ir * WARP_SIZE_GGUF + threadIdx.x;
       const int kbxd = kqs / QI8_1;
 
 #pragma unroll
@@ -73,7 +73,7 @@ static __device__ __forceinline__ void moe_q(
       }
 
       if (threadIdx.x < n_per_r / QK8_1) {
-        const int kby = threadIdx.x % (WARP_SIZE_GGUF / QI8_1);
+        const auto kby = threadIdx.x % (WARP_SIZE_GGUF / QI8_1);
         const int col_y_eff = token_offs[threadIdx.y] / top_k;
         const int block_x =
             ib0 * (qk / QK8_1) + ir * (WARP_SIZE_GGUF / QI8_1) + kby;
@@ -119,7 +119,7 @@ static __device__ __forceinline__ void moe_q(
 
 #pragma unroll
     for (int i = 0; i < mmq_y; i += WARP_SIZE_GGUF) {
-      const int row_dst = row_dst_0 + threadIdx.x + i;
+      const auto row_dst = row_dst_0 + threadIdx.x + i;
       if (row_dst >= nrows_dst) {
         continue;
       }
@@ -129,12 +129,12 @@ static __device__ __forceinline__ void moe_q(
 }
 
 #if defined(USE_ROCM)
-  #define MMQ_X_Q4_0 64
-  #define MMQ_Y_Q4_0 128
+  #define MOE_X_Q4_0 64
+  #define MOE_Y_Q4_0 128
   #define NWARPS_Q4_0 8
 #else
-  #define MMQ_X_Q4_0 4
-  #define MMQ_Y_Q4_0 32
+  #define MOE_X_Q4_0 4
+  #define MOE_Y_Q4_0 32
   #define NWARPS_Q4_0 4
 #endif
 
@@ -149,8 +149,8 @@ __launch_bounds__(WARP_SIZE_GGUF* NWARPS_Q4_0, 2)
              const int exp_stride, const int ncols_x, const int nrows_x,
              const int ncols_y, const int nrows_y, const int nrows_dst,
              const int top_k) {
-  const int mmq_x = MMQ_X_Q4_0;
-  const int mmq_y = MMQ_Y_Q4_0;
+  const int mmq_x = MOE_X_Q4_0;
+  const int mmq_y = MOE_Y_Q4_0;
   const int nwarps = NWARPS_Q4_0;
 
   moe_q<scalar_t, QK4_0, QR4_0, QI4_0, true, block_q4_0, mmq_x, mmq_y, nwarps,
@@ -167,8 +167,8 @@ static void ggml_moe_q4_0_q8_1_cuda(
     const int exp_stride, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, const int top_k,
     const int tokens_post_padded, cudaStream_t stream) {
-  int mmq_x = MMQ_X_Q4_0;
-  int mmq_y = MMQ_Y_Q4_0;
+  int mmq_x = MOE_X_Q4_0;
+  int mmq_y = MOE_Y_Q4_0;
   int nwarps = NWARPS_Q4_0;
 
   const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -190,12 +190,12 @@ static void ggml_moe_q4_0_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MMQ_X_Q4_1 64
-  #define MMQ_Y_Q4_1 128
+  #define MOE_X_Q4_1 64
+  #define MOE_Y_Q4_1 128
   #define NWARPS_Q4_1 8
 #else
-  #define MMQ_X_Q4_1 4
-  #define MMQ_Y_Q4_1 32
+  #define MOE_X_Q4_1 4
+  #define MOE_Y_Q4_1 32
   #define NWARPS_Q4_1 4
 #endif
 
@@ -210,8 +210,8 @@ __launch_bounds__(WARP_SIZE_GGUF* NWARPS_Q4_1, 2)
              const int exp_stride, const int ncols_x, const int nrows_x,
              const int ncols_y, const int nrows_y, const int nrows_dst,
              const int top_k) {
-  const int mmq_x = MMQ_X_Q4_1;
-  const int mmq_y = MMQ_Y_Q4_1;
+  const int mmq_x = MOE_X_Q4_1;
+  const int mmq_y = MOE_Y_Q4_1;
   const int nwarps = NWARPS_Q4_1;
 
   moe_q<scalar_t, QK4_1, QR4_1, QI4_1, true, block_q4_1, mmq_x, mmq_y, nwarps,
@@ -228,8 +228,8 @@ static void ggml_moe_q4_1_q8_1_cuda(
     const int exp_stride, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, const int top_k,
     const int tokens_post_padded, cudaStream_t stream) {
-  int mmq_x = MMQ_X_Q4_1;
-  int mmq_y = MMQ_Y_Q4_1;
+  int mmq_x = MOE_X_Q4_1;
+  int mmq_y = MOE_Y_Q4_1;
   int nwarps = NWARPS_Q4_1;
 
   const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -251,12 +251,12 @@ static void ggml_moe_q4_1_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MMQ_X_Q5_0 64
-  #define MMQ_Y_Q5_0 128
+  #define MOE_X_Q5_0 64
+  #define MOE_Y_Q5_0 128
   #define NWARPS_Q5_0 8
 #else
-  #define MMQ_X_Q5_0 4
-  #define MMQ_Y_Q5_0 32
+  #define MOE_X_Q5_0 4
+  #define MOE_Y_Q5_0 32
   #define NWARPS_Q5_0 4
 #endif
 
@@ -271,8 +271,8 @@ __launch_bounds__(WARP_SIZE_GGUF* NWARPS_Q5_0, 2)
              const int exp_stride, const int ncols_x, const int nrows_x,
              const int ncols_y, const int nrows_y, const int nrows_dst,
              const int top_k) {
-  const int mmq_x = MMQ_X_Q5_0;
-  const int mmq_y = MMQ_Y_Q5_0;
+  const int mmq_x = MOE_X_Q5_0;
+  const int mmq_y = MOE_Y_Q5_0;
   const int nwarps = NWARPS_Q5_0;
 
   moe_q<scalar_t, QK5_0, QR5_0, QI5_0, false, block_q5_0, mmq_x, mmq_y, nwarps,
@@ -289,8 +289,8 @@ static void ggml_moe_q5_0_q8_1_cuda(
     const int exp_stride, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, const int top_k,
     const int tokens_post_padded, cudaStream_t stream) {
-  const int mmq_x = MMQ_X_Q5_0;
-  const int mmq_y = MMQ_Y_Q5_0;
+  const int mmq_x = MOE_X_Q5_0;
+  const int mmq_y = MOE_Y_Q5_0;
   const int nwarps = NWARPS_Q5_0;
 
   const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -312,12 +312,12 @@ static void ggml_moe_q5_0_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MMQ_X_Q5_1 64
-  #define MMQ_Y_Q5_1 128
+  #define MOE_X_Q5_1 64
+  #define MOE_Y_Q5_1 128
   #define NWARPS_Q5_1 8
 #else
-  #define MMQ_X_Q5_1 4
-  #define MMQ_Y_Q5_1 32
+  #define MOE_X_Q5_1 4
+  #define MOE_Y_Q5_1 32
   #define NWARPS_Q5_1 4
 #endif
 
@@ -332,8 +332,8 @@ __launch_bounds__(WARP_SIZE_GGUF* NWARPS_Q5_1, 2)
              const int exp_stride, const int ncols_x, const int nrows_x,
              const int ncols_y, const int nrows_y, const int nrows_dst,
              const int top_k) {
-  const int mmq_x = MMQ_X_Q5_1;
-  const int mmq_y = MMQ_Y_Q5_1;
+  const int mmq_x = MOE_X_Q5_1;
+  const int mmq_y = MOE_Y_Q5_1;
   const int nwarps = NWARPS_Q5_1;
 
   moe_q<scalar_t, QK5_1, QR5_1, QI5_1, true, block_q5_1, mmq_x, mmq_y, nwarps,
@@ -350,8 +350,8 @@ static void ggml_moe_q5_1_q8_1_cuda(
     const int exp_stride, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, const int top_k,
     const int tokens_post_padded, cudaStream_t stream) {
-  const int mmq_x = MMQ_X_Q5_1;
-  const int mmq_y = MMQ_Y_Q5_1;
+  const int mmq_x = MOE_X_Q5_1;
+  const int mmq_y = MOE_Y_Q5_1;
   const int nwarps = NWARPS_Q5_1;
 
   const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -373,12 +373,12 @@ static void ggml_moe_q5_1_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MMQ_X_Q8_0 64
-  #define MMQ_Y_Q8_0 128
+  #define MOE_X_Q8_0 64
+  #define MOE_Y_Q8_0 128
   #define NWARPS_Q8_0 8
 #else
-  #define MMQ_X_Q8_0 4
-  #define MMQ_Y_Q8_0 32
+  #define MOE_X_Q8_0 4
+  #define MOE_Y_Q8_0 32
   #define NWARPS_Q8_0 4
 #endif
 
@@ -393,8 +393,8 @@ __launch_bounds__(WARP_SIZE_GGUF* NWARPS_Q8_0, 2)
              const int exp_stride, const int ncols_x, const int nrows_x,
              const int ncols_y, const int nrows_y, const int nrows_dst,
              const int top_k) {
-  const int mmq_x = MMQ_X_Q8_0;
-  const int mmq_y = MMQ_Y_Q8_0;
+  const int mmq_x = MOE_X_Q8_0;
+  const int mmq_y = MOE_Y_Q8_0;
   const int nwarps = NWARPS_Q8_0;
 
   moe_q<scalar_t, QK8_0, QR8_0, QI8_0, false, block_q8_0, mmq_x, mmq_y, nwarps,
@@ -411,8 +411,8 @@ static void ggml_moe_q8_0_q8_1_cuda(
     const int exp_stride, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, const int top_k,
     const int tokens_post_padded, cudaStream_t stream) {
-  const int mmq_x = MMQ_X_Q8_0;
-  const int mmq_y = MMQ_Y_Q8_0;
+  const int mmq_x = MOE_X_Q8_0;
+  const int mmq_y = MOE_Y_Q8_0;
   const int nwarps = NWARPS_Q8_0;
 
   const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -434,12 +434,12 @@ static void ggml_moe_q8_0_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MMQ_X_Q2_K 64
-  #define MMQ_Y_Q2_K 128
+  #define MOE_X_Q2_K 64
+  #define MOE_Y_Q2_K 128
   #define NWARPS_Q2_K 8
 #else
-  #define MMQ_X_Q2_K 4
-  #define MMQ_Y_Q2_K 32
+  #define MOE_X_Q2_K 4
+  #define MOE_Y_Q2_K 32
   #define NWARPS_Q2_K 4
 #endif
 
@@ -454,8 +454,8 @@ __launch_bounds__(WARP_SIZE_GGUF* NWARPS_Q2_K, 2)
              const int exp_stride, const int ncols_x, const int nrows_x,
              const int ncols_y, const int nrows_y, const int nrows_dst,
              const int top_k) {
-  const int mmq_x = MMQ_X_Q2_K;
-  const int mmq_y = MMQ_Y_Q2_K;
+  const int mmq_x = MOE_X_Q2_K;
+  const int mmq_y = MOE_Y_Q2_K;
   const int nwarps = NWARPS_Q2_K;
 
   moe_q<scalar_t, QK_K, QR2_K, QI2_K, false, block_q2_K, mmq_x, mmq_y, nwarps,
@@ -472,8 +472,8 @@ static void ggml_moe_q2_K_q8_1_cuda(
     const int exp_stride, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, const int top_k,
     const int tokens_post_padded, cudaStream_t stream) {
-  const int mmq_x = MMQ_X_Q2_K;
-  const int mmq_y = MMQ_Y_Q2_K;
+  const int mmq_x = MOE_X_Q2_K;
+  const int mmq_y = MOE_Y_Q2_K;
   const int nwarps = NWARPS_Q2_K;
 
   const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -495,12 +495,12 @@ static void ggml_moe_q2_K_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MMQ_X_Q3_K 64
-  #define MMQ_Y_Q3_K 128
+  #define MOE_X_Q3_K 64
+  #define MOE_Y_Q3_K 128
   #define NWARPS_Q3_K 8
 #else
-  #define MMQ_X_Q3_K 4
-  #define MMQ_Y_Q3_K 32
+  #define MOE_X_Q3_K 4
+  #define MOE_Y_Q3_K 32
   #define NWARPS_Q3_K 4
 #endif
 
@@ -516,8 +516,8 @@ __launch_bounds__(WARP_SIZE_GGUF* NWARPS_Q3_K, 2)
              const int ncols_y, const int nrows_y, const int nrows_dst,
              const int top_k) {
 
-  const int mmq_x = MMQ_X_Q3_K;
-  const int mmq_y = MMQ_Y_Q3_K;
+  const int mmq_x = MOE_X_Q3_K;
+  const int mmq_y = MOE_Y_Q3_K;
   const int nwarps = NWARPS_Q3_K;
 
   moe_q<scalar_t, QK_K, QR3_K, QI3_K, false, block_q3_K, mmq_x, mmq_y, nwarps,
@@ -533,8 +533,8 @@ static void ggml_moe_q3_K_q8_1_cuda(
     const int exp_stride, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, const int top_k,
     const int tokens_post_padded, cudaStream_t stream) {
-  const int mmq_x = MMQ_X_Q3_K;
-  const int mmq_y = MMQ_Y_Q3_K;
+  const int mmq_x = MOE_X_Q3_K;
+  const int mmq_y = MOE_Y_Q3_K;
   const int nwarps = NWARPS_Q3_K;
 
   const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -556,12 +556,12 @@ static void ggml_moe_q3_K_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MMQ_X_Q4_K 64
-  #define MMQ_Y_Q4_K 128
+  #define MOE_X_Q4_K 64
+  #define MOE_Y_Q4_K 128
   #define NWARPS_Q4_K 8
 #else
-  #define MMQ_X_Q4_K 4
-  #define MMQ_Y_Q4_K 32
+  #define MOE_X_Q4_K 4
+  #define MOE_Y_Q4_K 32
   #define NWARPS_Q4_K 4
 #endif
 
@@ -576,8 +576,8 @@ __launch_bounds__(WARP_SIZE_GGUF* NWARPS_Q4_K, 2)
              const int exp_stride, const int ncols_x, const int nrows_x,
              const int ncols_y, const int nrows_y, const int nrows_dst,
              const int top_k) {
-  const int mmq_x = MMQ_X_Q4_K;
-  const int mmq_y = MMQ_Y_Q4_K;
+  const int mmq_x = MOE_X_Q4_K;
+  const int mmq_y = MOE_Y_Q4_K;
   const int nwarps = NWARPS_Q4_K;
 
   moe_q<scalar_t, QK_K, QR4_K, QI4_K, true, block_q4_K, mmq_x, mmq_y, nwarps,
@@ -594,8 +594,8 @@ static void ggml_moe_q4_K_q8_1_cuda(
     const int exp_stride, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, const int top_k,
     const int tokens_post_padded, cudaStream_t stream) {
-  const int mmq_x = MMQ_X_Q4_K;
-  const int mmq_y = MMQ_Y_Q4_K;
+  const int mmq_x = MOE_X_Q4_K;
+  const int mmq_y = MOE_Y_Q4_K;
   const int nwarps = NWARPS_Q4_K;
 
   const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -617,12 +617,12 @@ static void ggml_moe_q4_K_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MMQ_X_Q5_K 64
-  #define MMQ_Y_Q5_K 128
+  #define MOE_X_Q5_K 64
+  #define MOE_Y_Q5_K 128
   #define NWARPS_Q5_K 8
 #else
-  #define MMQ_X_Q5_K 4
-  #define MMQ_Y_Q5_K 32
+  #define MOE_X_Q5_K 4
+  #define MOE_Y_Q5_K 32
   #define NWARPS_Q5_K 4
 #endif
 
@@ -637,8 +637,8 @@ __launch_bounds__(WARP_SIZE_GGUF* NWARPS_Q5_K, 2)
              const int exp_stride, const int ncols_x, const int nrows_x,
              const int ncols_y, const int nrows_y, const int nrows_dst,
              const int top_k) {
-  const int mmq_x = MMQ_X_Q5_K;
-  const int mmq_y = MMQ_Y_Q5_K;
+  const int mmq_x = MOE_X_Q5_K;
+  const int mmq_y = MOE_Y_Q5_K;
   const int nwarps = NWARPS_Q5_K;
 
   moe_q<scalar_t, QK_K, QR5_K, QI5_K, true, block_q5_K, mmq_x, mmq_y, nwarps,
@@ -655,8 +655,8 @@ static void ggml_moe_q5_K_q8_1_cuda(
     const int exp_stride, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, const int top_k,
     const int tokens_post_padded, cudaStream_t stream) {
-  const int mmq_x = MMQ_X_Q5_K;
-  const int mmq_y = MMQ_Y_Q5_K;
+  const int mmq_x = MOE_X_Q5_K;
+  const int mmq_y = MOE_Y_Q5_K;
   const int nwarps = NWARPS_Q5_K;
 
   const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
@@ -678,12 +678,12 @@ static void ggml_moe_q5_K_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MMQ_X_Q6_K 64
-  #define MMQ_Y_Q6_K 128
+  #define MOE_X_Q6_K 64
+  #define MOE_Y_Q6_K 128
   #define NWARPS_Q6_K 8
 #else
-  #define MMQ_X_Q6_K 4
-  #define MMQ_Y_Q6_K 32
+  #define MOE_X_Q6_K 4
+  #define MOE_Y_Q6_K 32
   #define NWARPS_Q6_K 4
 #endif
 
@@ -698,8 +698,8 @@ __launch_bounds__(WARP_SIZE_GGUF* NWARPS_Q6_K, 2)
              const int exp_stride, const int ncols_x, const int nrows_x,
              const int ncols_y, const int nrows_y, const int nrows_dst,
              const int top_k) {
-  const int mmq_x = MMQ_X_Q6_K;
-  const int mmq_y = MMQ_Y_Q6_K;
+  const int mmq_x = MOE_X_Q6_K;
+  const int mmq_y = MOE_Y_Q6_K;
   const int nwarps = NWARPS_Q6_K;
 
   moe_q<scalar_t, QK_K, QR6_K, QI6_K, false, block_q6_K, mmq_x, mmq_y, nwarps,
@@ -716,8 +716,8 @@ static void ggml_moe_q6_K_q8_1_cuda(
     const int exp_stride, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, const int top_k,
     const int tokens_post_padded, cudaStream_t stream) {
-  const int mmq_x = MMQ_X_Q6_K;
-  const int mmq_y = MMQ_Y_Q6_K;
+  const int mmq_x = MOE_X_Q6_K;
+  const int mmq_y = MOE_Y_Q6_K;
   const int nwarps = NWARPS_Q6_K;
 
   const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;

csrc/quantization/gptq/q_gemm.cu

@@ -199,12 +199,12 @@ __global__ void gemm_half_q_half_gptq_4bit_kernel(
   MatrixView_q4_row b_gptq_qzeros_(b_gptq_qzeros, groups, size_n);
   MatrixView_half b_gptq_scales_(b_gptq_scales, groups, size_n);
 
-  int t = threadIdx.x;
+  auto t = threadIdx.x;
 
   // Block
-  int offset_n = blockIdx.x * BLOCK_KN_SIZE * 4;
-  int offset_m = blockIdx.y * m_count;
-  int offset_k = blockIdx.z * BLOCK_KN_SIZE;
+  auto offset_n = blockIdx.x * BLOCK_KN_SIZE * 4;
+  auto offset_m = blockIdx.y * m_count;
+  auto offset_k = blockIdx.z * BLOCK_KN_SIZE;
 
   [[maybe_unused]] int end_n = min(offset_n + BLOCK_KN_SIZE * 4, size_n);
   [[maybe_unused]] int end_m = min(offset_m + m_count, size_m);
@@ -337,12 +337,12 @@ __global__ void gemm_half_q_half_gptq_2bit_kernel(
   MatrixView_q2_row b_gptq_qzeros_(b_gptq_qzeros, groups, size_n);
   MatrixView_half b_gptq_scales_(b_gptq_scales, groups, size_n);
 
-  int t = threadIdx.x;
+  auto t = threadIdx.x;
 
   // Block
-  int offset_n = blockIdx.x * BLOCK_KN_SIZE * 4;
-  int offset_m = blockIdx.y * m_count;
-  int offset_k = blockIdx.z * BLOCK_KN_SIZE;
+  auto offset_n = blockIdx.x * BLOCK_KN_SIZE * 4;
+  auto offset_m = blockIdx.y * m_count;
+  auto offset_k = blockIdx.z * BLOCK_KN_SIZE;
 
   [[maybe_unused]] int end_n = min(offset_n + BLOCK_KN_SIZE * 4, size_n);
   [[maybe_unused]] int end_m = min(offset_m + m_count, size_m);
@@ -458,12 +458,12 @@ __global__ void gemm_half_q_half_gptq_3bit_kernel(
   MatrixView_q3_row b_gptq_qzeros_(b_gptq_qzeros, groups, size_n);
   MatrixView_half b_gptq_scales_(b_gptq_scales, groups, size_n);
 
-  int t = threadIdx.x;
+  auto t = threadIdx.x;
 
   // Block
-  int offset_n = blockIdx.x * BLOCK_KN_SIZE * 4;
-  int offset_m = blockIdx.y * m_count;
-  int offset_k = blockIdx.z * BLOCK_KN_SIZE;
+  auto offset_n = blockIdx.x * BLOCK_KN_SIZE * 4;
+  auto offset_m = blockIdx.y * m_count;
+  auto offset_k = blockIdx.z * BLOCK_KN_SIZE;
 
   [[maybe_unused]] int end_n = min(offset_n + BLOCK_KN_SIZE * 4, size_n);
   [[maybe_unused]] int end_m = min(offset_m + m_count, size_m);
@@ -586,12 +586,12 @@ __global__ void gemm_half_q_half_gptq_8bit_kernel(
   MatrixView_q8_row b_gptq_qzeros_(b_gptq_qzeros, groups, size_n);
   MatrixView_half b_gptq_scales_(b_gptq_scales, groups, size_n);
 
-  int t = threadIdx.x;
+  auto t = threadIdx.x;
 
   // Block
-  int offset_n = blockIdx.x * BLOCK_KN_SIZE * 4;
-  int offset_m = blockIdx.y * m_count;
-  int offset_k = blockIdx.z * BLOCK_KN_SIZE;
+  auto offset_n = blockIdx.x * BLOCK_KN_SIZE * 4;
+  auto offset_m = blockIdx.y * m_count;
+  auto offset_k = blockIdx.z * BLOCK_KN_SIZE;
 
   [[maybe_unused]] int end_n = min(offset_n + BLOCK_KN_SIZE * 4, size_n);
   [[maybe_unused]] int end_m = min(offset_m + m_count, size_m);
@@ -765,14 +765,14 @@ __global__ void reconstruct_exllama_8bit_kernel(
   MatrixView_q8_row b_gptq_qzeros_(b_gptq_qzeros, groups, size_n);
   MatrixView_half b_gptq_scales_(b_gptq_scales, groups, size_n);
 
-  int offset_k = BLOCK_KN_SIZE * blockIdx.y;
-  int offset_n = BLOCK_KN_SIZE * blockIdx.x * 4;
+  auto offset_k = BLOCK_KN_SIZE * blockIdx.y;
+  auto offset_n = BLOCK_KN_SIZE * blockIdx.x * 4;
 
   int end_k = min(offset_k + BLOCK_KN_SIZE, size_k);
 
   // Preload remapping table
   __shared__ int perm[BLOCK_KN_SIZE];
-  int t = threadIdx.x;
+  auto t = threadIdx.x;
 
   if (b_q_perm) {
     if (offset_k + t < size_k) perm[t] = b_q_perm[offset_k + t];
@@ -862,14 +862,14 @@ __global__ void reconstruct_exllama_4bit_kernel(
   MatrixView_q4_row b_gptq_qzeros_(b_gptq_qzeros, groups, size_n);
   MatrixView_half b_gptq_scales_(b_gptq_scales, groups, size_n);
 
-  int offset_k = BLOCK_KN_SIZE * blockIdx.y;
-  int offset_n = BLOCK_KN_SIZE * blockIdx.x * 4;
+  auto offset_k = BLOCK_KN_SIZE * blockIdx.y;
+  auto offset_n = BLOCK_KN_SIZE * blockIdx.x * 4;
 
   int end_k = min(offset_k + BLOCK_KN_SIZE, size_k);
 
   // Preload remapping table
   __shared__ int perm[BLOCK_KN_SIZE];
-  int t = threadIdx.x;
+  auto t = threadIdx.x;
 
   if (b_q_perm) {
     if (offset_k + t < size_k) perm[t] = b_q_perm[offset_k + t];
@@ -967,14 +967,14 @@ __global__ void reconstruct_exllama_3bit_kernel(
   MatrixView_q3_row b_gptq_qzeros_(b_gptq_qzeros, groups, size_n);
   MatrixView_half b_gptq_scales_(b_gptq_scales, groups, size_n);
 
-  int offset_k = BLOCK_KN_SIZE * blockIdx.y;
-  int offset_n = BLOCK_KN_SIZE * blockIdx.x * 4;
+  auto offset_k = BLOCK_KN_SIZE * blockIdx.y;
+  auto offset_n = BLOCK_KN_SIZE * blockIdx.x * 4;
 
   int end_k = min(offset_k + BLOCK_KN_SIZE, size_k);
 
   // Preload remapping table
   __shared__ int perm[BLOCK_KN_SIZE];
-  int t = threadIdx.x;
+  auto t = threadIdx.x;
 
   if (b_q_perm) {
     if (offset_k + t < size_k) perm[t] = b_q_perm[offset_k + t];
@@ -1065,14 +1065,14 @@ __global__ void reconstruct_exllama_2bit_kernel(
   MatrixView_q2_row b_gptq_qzeros_(b_gptq_qzeros, groups, size_n);
   MatrixView_half b_gptq_scales_(b_gptq_scales, groups, size_n);
 
-  int offset_k = BLOCK_KN_SIZE * blockIdx.y;
-  int offset_n = BLOCK_KN_SIZE * blockIdx.x * 4;
+  auto offset_k = BLOCK_KN_SIZE * blockIdx.y;
+  auto offset_n = BLOCK_KN_SIZE * blockIdx.x * 4;
 
   int end_k = min(offset_k + BLOCK_KN_SIZE, size_k);
 
   // Preload remapping table
   __shared__ int perm[BLOCK_KN_SIZE];
-  int t = threadIdx.x;
+  auto t = threadIdx.x;
 
   if (b_q_perm) {
     if (offset_k + t < size_k) perm[t] = b_q_perm[offset_k + t];
@@ -1181,11 +1181,11 @@ __global__ void gemm_half_q_half_alt_4bit_kernel(
   int zero_width = width / 8;
   int vec_height = height * 4;
   const int blockwidth2 = BLOCK_KN_SIZE / 2;
-  int b = blockIdx.y * BLOCK_M_SIZE_MAX;
+  auto b = blockIdx.y * BLOCK_M_SIZE_MAX;
   int b_end = min(BLOCK_M_SIZE_MAX, batch - b);
-  int h = BLOCK_KN_SIZE * blockIdx.z / 8;
+  auto h = BLOCK_KN_SIZE * blockIdx.z / 8;
   int h_end = min(BLOCK_KN_SIZE / 8, height - h) * 4;
-  int w = BLOCK_KN_SIZE * blockIdx.x + threadIdx.x;
+  auto w = BLOCK_KN_SIZE * blockIdx.x + threadIdx.x;
 
   __shared__ half2 blockvec[BLOCK_M_SIZE_MAX][blockwidth2];
   if (threadIdx.x < h_end) {
@@ -1197,8 +1197,8 @@ __global__ void gemm_half_q_half_alt_4bit_kernel(
   }
 
   __shared__ half2 deq2[256][8];
-  int val = threadIdx.x / 8;
-  int off = threadIdx.x % 8;
+  auto val = threadIdx.x / 8;
+  auto off = threadIdx.x % 8;
   for (; val < 256; val += BLOCK_KN_SIZE / 8) {
     deq2[val][off] =
         __halves2half2(__int2half_rn(val & 0xF), __int2half_rn(val >> 4));
@@ -1280,11 +1280,11 @@ __global__ void gemm_half_q_half_alt_8bit_kernel(
   int zero_width = width / 4;
   int vec_height = height * 2;
   const int blockwidth2 = BLOCK_KN_SIZE / 2;
-  int b = blockIdx.y * BLOCK_M_SIZE_MAX;
+  auto b = blockIdx.y * BLOCK_M_SIZE_MAX;
   int b_end = min(BLOCK_M_SIZE_MAX, batch - b);
-  int h = BLOCK_KN_SIZE * blockIdx.z / 4;
+  auto h = BLOCK_KN_SIZE * blockIdx.z / 4;
   int h_end = min(BLOCK_KN_SIZE / 4, height - h) * 2;
-  int w = BLOCK_KN_SIZE * blockIdx.x + threadIdx.x;
+  auto w = BLOCK_KN_SIZE * blockIdx.x + threadIdx.x;
 
   __shared__ half2 blockvec[BLOCK_M_SIZE_MAX][blockwidth2];
   if (threadIdx.x < h_end) {
@@ -1393,8 +1393,8 @@ __global__ void reconstruct_gptq_kernel(const uint32_t* __restrict__ w,
                                         half* __restrict__ out) {
   // Start of block
 
-  int column = BLOCK_KN_SIZE * blockIdx.x + threadIdx.x;
-  int row = blockIdx.y * 32 / bit;
+  auto column = BLOCK_KN_SIZE * blockIdx.x + threadIdx.x;
+  auto row = blockIdx.y * 32 / bit;
   if (column >= width) return;
 
   // Views
@@ -1425,8 +1425,8 @@ __global__ void reconstruct_gptq_3bit_kernel(
     const int height, const int width, const int group,
     half* __restrict__ out) {
   // Start of block
-  int column = BLOCK_KN_SIZE * blockIdx.x + threadIdx.x;
-  int row = blockIdx.y * 32;
+  auto column = BLOCK_KN_SIZE * blockIdx.x + threadIdx.x;
+  auto row = blockIdx.y * 32;
   if (column >= width) return;
 
   // Views
@@ -1542,7 +1542,7 @@ void gemm_half_q_half_cuda(cublasHandle_t cublas_handle, const half* a,
 
 __global__ void shuffle_4bit_kernel(uint32_t* __restrict__ b_q_weight,
                                     const int size_k, const int size_n) {
-  int n = blockIdx.x * THREADS_X + threadIdx.x;
+  auto n = blockIdx.x * THREADS_X + threadIdx.x;
   if (n >= size_n) return;
   int k = 0;
   uint32_t* b_ptr = b_q_weight + n;
@@ -1555,7 +1555,7 @@ __global__ void shuffle_4bit_kernel(uint32_t* __restrict__ b_q_weight,
 
 __global__ void shuffle_8bit_kernel(uint32_t* __restrict__ b_q_weight,
                                     const int size_k, const int size_n) {
-  int n = blockIdx.x * THREADS_X + threadIdx.x;
+  auto n = blockIdx.x * THREADS_X + threadIdx.x;
   if (n >= size_n) return;
   int k = 0;
   uint32_t* b_ptr = b_q_weight + n;
@@ -1568,7 +1568,7 @@ __global__ void shuffle_8bit_kernel(uint32_t* __restrict__ b_q_weight,
 
 __global__ void shuffle_2bit_kernel(uint32_t* __restrict__ b_q_weight,
                                     const int size_k, const int size_n) {
-  int n = blockIdx.x * THREADS_X + threadIdx.x;
+  auto n = blockIdx.x * THREADS_X + threadIdx.x;
   if (n >= size_n) return;
   int k = 0;
   uint32_t* b_ptr = b_q_weight + n;
@@ -1581,7 +1581,7 @@ __global__ void shuffle_2bit_kernel(uint32_t* __restrict__ b_q_weight,
 
 __global__ void shuffle_3bit_kernel(uint32_t* __restrict__ b_q_weight,
                                     const int size_k, const int size_n) {
-  int n = blockIdx.x * THREADS_X + threadIdx.x;
+  auto n = blockIdx.x * THREADS_X + threadIdx.x;
   if (n >= size_n) return;
   int k = 0;
   uint32_t* b_ptr = b_q_weight + n;
@@ -1599,9 +1599,9 @@ __global__ void make_sequential_4bit_kernel(const uint32_t* __restrict__ w,
   const uint64_t* w2 = (uint64_t*)w;
   uint64_t* w_new2 = (uint64_t*)w_new;
   int w2_stride = w_width >> 1;
-  int w2_column = THREADS_X * blockIdx.x + threadIdx.x;
+  auto w2_column = THREADS_X * blockIdx.x + threadIdx.x;
   if (w2_column >= w2_stride) return;
-  int w_new2_row = blockIdx.y;
+  auto w_new2_row = blockIdx.y;
   int q_perm_idx = w_new2_row << 3;
   uint64_t dst = 0;
 
@@ -1630,9 +1630,9 @@ __global__ void make_sequential_2bit_kernel(const uint32_t* __restrict__ w,
   const uint64_t* w2 = (uint64_t*)w;
   uint64_t* w_new2 = (uint64_t*)w_new;
   int w2_stride = w_width >> 1;
-  int w2_column = THREADS_X * blockIdx.x + threadIdx.x;
+  auto w2_column = THREADS_X * blockIdx.x + threadIdx.x;
   if (w2_column >= w2_stride) return;
-  int w_new2_row = blockIdx.y;
+  auto w_new2_row = blockIdx.y;
   int q_perm_idx = w_new2_row << 4;
   uint64_t dst = 0;
 
@@ -1658,10 +1658,10 @@ __global__ void make_sequential_3bit_kernel(const uint32_t* __restrict__ w,
                                             uint32_t* __restrict__ w_new,
                                             const int* __restrict__ q_perm,
                                             const int w_width) {
-  int w_column = THREADS_X * blockIdx.x + threadIdx.x;
+  auto w_column = THREADS_X * blockIdx.x + threadIdx.x;
   if (w_column >= w_width) return;
-  int w_new_row = blockIdx.y * 3;
-  int q_perm_idx = blockIdx.y << 5;
+  auto w_new_row = blockIdx.y * 3;
+  auto q_perm_idx = blockIdx.y << 5;
   uint32_t dst[3] = {0, 0, 0};
 
 #pragma unroll
@@ -1744,9 +1744,9 @@ __global__ void make_sequential_8bit_kernel(const uint32_t* __restrict__ w,
   const uint64_t* w2 = (uint64_t*)w;
   uint64_t* w_new2 = (uint64_t*)w_new;
   int w2_stride = w_width >> 1;
-  int w2_column = THREADS_X * blockIdx.x + threadIdx.x;
+  auto w2_column = THREADS_X * blockIdx.x + threadIdx.x;
   if (w2_column >= w2_stride) return;
-  int w_new2_row = blockIdx.y;
+  auto w_new2_row = blockIdx.y;
   int q_perm_idx = w_new2_row << 2;
   uint64_t dst = 0;
 

csrc/quantization/gptq_allspark/allspark_qgemm_w8a16.cu

@@ -55,11 +55,11 @@ struct GmemTile_W8A16_PerC_MtilexNtilex32_multistage_SM8x_SplitK {
     this_block_B_base_ptr = params.B_ptr + blockIdx.y * Ntile * params.K +
                             blockIdx.z * params.SplitK * 4;
 
-    const int lane_id = threadIdx.x % WARP_SIZE;
+    const auto lane_id = threadIdx.x % WARP_SIZE;
 
     // For matrix A, a block load/store Mtile(row) x 32(col) elements in
     // multiple iters, 8x4 warp load/store 8(row) x 32(col) elements per iter
-    const int Aldg_row_base_idx = threadIdx.x / 4;
+    const auto Aldg_row_base_idx = threadIdx.x / 4;
     Aldg_col_idx = (threadIdx.x % 4) * LDG_ELEMENT_CNT_A;
     const int Aldg_base_offset = Aldg_row_base_idx * params.K + Aldg_col_idx;
 
@@ -67,7 +67,7 @@ struct GmemTile_W8A16_PerC_MtilexNtilex32_multistage_SM8x_SplitK {
     // elements of N32K16 packing in multiple iters, 4x8 warp load/store 4(row)
     // * 128(col) per iter
     Bldg_col_idx = (threadIdx.x % 8) * LDG_ELEMENT_CNT_B;
-    const int Bldg_row_base_idx = threadIdx.x / 8;
+    const auto Bldg_row_base_idx = threadIdx.x / 8;
     const int Bldg_base_offset =
         Bldg_row_base_idx * params.K * 4 + Bldg_col_idx;
 
@@ -89,7 +89,7 @@ struct GmemTile_W8A16_PerC_MtilexNtilex32_multistage_SM8x_SplitK {
     B_ldg_guard = 0;
   #pragma unroll
     for (int i = 0; i < (Mtile + M_SIZE_ONE_LOAD - 1) / M_SIZE_ONE_LOAD; ++i) {
-      int m_idx = blockIdx.x * Mtile + Aldg_row_base_idx + i * M_SIZE_ONE_LOAD;
+      auto m_idx = blockIdx.x * Mtile + Aldg_row_base_idx + i * M_SIZE_ONE_LOAD;
       if (m_idx < params.M) {
         A_ldg_guard |= (1u << i);
       }
@@ -98,8 +98,8 @@ struct GmemTile_W8A16_PerC_MtilexNtilex32_multistage_SM8x_SplitK {
     const int N_padded = (params.N + 31) / 32 * 32;
   #pragma unroll
     for (int i = 0; i < (Ntile + N_SIZE_ONE_LOAD - 1) / N_SIZE_ONE_LOAD; ++i) {
-      int n_idx = blockIdx.y * Ntile + (Bldg_row_base_idx / 8) * 32 +
-                  i * N_SIZE_ONE_LOAD;
+      auto n_idx = blockIdx.y * Ntile + (Bldg_row_base_idx / 8) * 32 +
+                   i * N_SIZE_ONE_LOAD;
       if (n_idx < N_padded) {
         B_ldg_guard |= (1u << i);
       }
@@ -355,7 +355,7 @@ struct ComputeTile_W8A16_PerC_MtilexNtilex32_multistage_SM8x_SplitK {
   __device__ void fused_splitk_reduce() {
     // need splitk-reduce if enable splitk
     if (gridDim.z > 1) {
-      int blk_red_idx = blockIdx.x * gridDim.y + blockIdx.y;
+      auto blk_red_idx = blockIdx.x * gridDim.y + blockIdx.y;
       // Wait for all previous blocks in the splitk direction to accumulate the
       // results into C_tmp
       if (threadIdx.x == 0) {
@@ -371,7 +371,7 @@ struct ComputeTile_W8A16_PerC_MtilexNtilex32_multistage_SM8x_SplitK {
       }
       __syncthreads();
 
-      int C_tmp_base_offset = blk_red_idx * Mtile * Ntile + threadIdx.x * 4;
+      auto C_tmp_base_offset = blk_red_idx * Mtile * Ntile + threadIdx.x * 4;
       if (blockIdx.z != 0) {
         // expecting that temporary register here reuses the previous A&B frag
         // register
@@ -456,7 +456,7 @@ struct ComputeTile_W8A16_PerC_MtilexNtilex32_multistage_SM8x_SplitK {
 
     FType* C_base_ptr = this_block_C_base_ptr + store_c_base_offset;
     // C_tile lds and stg
-    int m_base_idx = store_c_row_base_idx + blockIdx.x * Mtile;
+    auto m_base_idx = store_c_row_base_idx + blockIdx.x * Mtile;
     bool n_guard = (store_c_col_idx + blockIdx.y * Ntile) < params.N;
     if (WARP_NTILE == 32) {
       int lds_c_base_offset = warp_id * Mtile * WARP_NTILE +
@@ -580,9 +580,9 @@ __global__ void __launch_bounds__(BLOCK)
   int sts_stage_idx = 0;
   int lds_stage_idx = 0;
 
-  int tb_k_slice = blockIdx.z * params.SplitK + params.SplitK <= params.K
-                       ? params.SplitK
-                       : params.K - blockIdx.z * params.SplitK;
+  auto tb_k_slice = blockIdx.z * params.SplitK + params.SplitK <= params.K
+                        ? params.SplitK
+                        : params.K - blockIdx.z * params.SplitK;
   int k_tiles = (tb_k_slice + 31) / 32;
   int first_k_tile = tb_k_slice - (k_tiles - 1) * 32;
 
@@ -777,13 +777,13 @@ __global__ void restore_N32_K16_dequantize_rhs_w8a16_perc_kernel(
     const QT* qdata, const FT* scales, const FT* zeros, FT* fdata,
     const int N_32align, const int N, const int K) {
   __shared__ FT smem[64 * 32];
-  int warp_id = threadIdx.x / 32;
-  int lane_id = threadIdx.x % 32;
-  const int src_row_idx = blockIdx.x * 8 + lane_id / 4;
+  auto warp_id = threadIdx.x / 32;
+  auto lane_id = threadIdx.x % 32;
+  const auto src_row_idx = blockIdx.x * 8 + lane_id / 4;
   const int src_col_idx =
       blockIdx.y * 64 * 4 + warp_id * 16 * 4 + (lane_id % 4) * 16;
   const int src_offset = src_row_idx * K * 4 + src_col_idx;
-  int params_nidx = blockIdx.x * 32 + (lane_id / 4) * 4;
+  auto params_nidx = blockIdx.x * 32 + (lane_id / 4) * 4;
 
   QT qval_reg[16];
   const QT* pdata = qdata + src_offset;
@@ -829,8 +829,8 @@ __global__ void restore_N32_K16_dequantize_rhs_w8a16_perc_kernel(
         *reinterpret_cast<uint4*>(smem + lds_base_offset + i * 32 * 32);
   }
 
-  const int dst_row_base_kidx = blockIdx.y * 64 + threadIdx.x / 4;
-  const int dst_col_nidx = blockIdx.x * 32 + (threadIdx.x % 4) * 8;
+  const auto dst_row_base_kidx = blockIdx.y * 64 + threadIdx.x / 4;
+  const auto dst_col_nidx = blockIdx.x * 32 + (threadIdx.x % 4) * 8;
   #pragma unroll
   for (int i = 0; i < 2; ++i) {
     int dst_row_kidx = dst_row_base_kidx + i * 32;
@@ -1008,4 +1008,4 @@ torch::Tensor allspark_w8a16_gemm(
 
 TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
   m.impl("allspark_w8a16_gemm", &allspark_w8a16_gemm);
-}
+}

csrc/quantization/gptq_allspark/allspark_repack.cu

@@ -13,8 +13,8 @@ __global__ void __launch_bounds__(128)
         const uint8_t* B, const FType* B_scale, const FType* B_zero,
         uint8_t* B_result, FType* B_scale_result, FType* B_zero_result,
         const int K, const int N, const int N_32align) {
-  const int lane_id = threadIdx.x % 32;
-  const int warp_id = threadIdx.x / 32;
+  const auto lane_id = threadIdx.x % 32;
+  const auto warp_id = threadIdx.x / 32;
 
   if (blockIdx.x != gridDim.x - 1) {
     // Load B
@@ -50,7 +50,7 @@ __global__ void __launch_bounds__(128)
     }
 
     // Store B
-    const int dst_row_base_idx = blockIdx.y * (128 / 4) + (lane_id / 8) * 8;
+    const auto dst_row_base_idx = blockIdx.y * (128 / 4) + (lane_id / 8) * 8;
     const int dst_col_idx =
         blockIdx.x * (64 * 4) + warp_id * 64 + (lane_id % 8) * 8;
     for (int i = 0; i < 8; ++i) {
@@ -65,7 +65,7 @@ __global__ void __launch_bounds__(128)
   } else {
     // Load B_scale and B_zero
     FType b_scale_reg, b_zero_reg;
-    int src_offset = blockIdx.y * 128 + threadIdx.x;
+    auto src_offset = blockIdx.y * 128 + threadIdx.x;
     ldg16_cg_0(b_scale_reg, B_scale + src_offset, src_offset < N);
     if (B_zero != nullptr)
       ldg16_cg_0(b_zero_reg, B_zero + src_offset, src_offset < N);

csrc/quantization/gptq_allspark/allspark_utils.cuh

@@ -62,7 +62,7 @@ template <typename FType, int BLOCK, int N_MATRIX>
 __global__ void f16_gemm_splitk_reduce_kernel(const FType* C_split, FType* C,
                                               uint32_t n, uint32_t n_matrix,
                                               uint32_t matrix_size) {
-  int idx = blockIdx.x * BLOCK + threadIdx.x;
+  auto idx = blockIdx.x * BLOCK + threadIdx.x;
 
   if (idx >= matrix_size) {
     return;
@@ -407,4 +407,4 @@ static __device__ half2 inline num2num2(const half x) {
   return __half2half2(x);
 }
 
-}  // namespace allspark
+}  // namespace allspark

csrc/quantization/gptq_marlin/awq_marlin_repack.cu

@@ -14,7 +14,7 @@ __global__ void awq_marlin_repack_kernel(
   int n_tiles = size_n / tile_n_size;
   int block_k_tiles = div_ceil(k_tiles, gridDim.x);
 
-  int start_k_tile = blockIdx.x * block_k_tiles;
+  auto start_k_tile = blockIdx.x * block_k_tiles;
   if (start_k_tile >= k_tiles) {
     return;
   }
@@ -51,8 +51,8 @@ __global__ void awq_marlin_repack_kernel(
     int4* sh_ptr = sh + stage_size * pipe;
 
     if (threadIdx.x < stage_size) {
-      int k_id = threadIdx.x / stage_n_threads;
-      int n_id = threadIdx.x % stage_n_threads;
+      auto k_id = threadIdx.x / stage_n_threads;
+      auto n_id = threadIdx.x % stage_n_threads;
 
       int first_k = k_tile_id * tile_k_size;
 
@@ -70,8 +70,8 @@ __global__ void awq_marlin_repack_kernel(
       return;
     }
 
-    int warp_id = threadIdx.x / 32;
-    int th_id = threadIdx.x % 32;
+    auto warp_id = threadIdx.x / 32;
+    auto th_id = threadIdx.x % 32;
 
     if (warp_id >= 4) {
       return;
@@ -265,4 +265,4 @@ TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
 
 TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, Meta, m) {
   m.impl("awq_marlin_repack", &awq_marlin_repack_meta);
-}
+}

csrc/quantization/gptq_marlin/gptq_marlin.cu

@@ -42,7 +42,7 @@ namespace marlin {
 __global__ void permute_cols_kernel(int4 const* __restrict__ a_int4_ptr,
                                     int const* __restrict__ perm_int_ptr,
                                     int4* __restrict__ out_int4_ptr, int size_m,
-                                    int size_k, int block_rows) {}
+                                    int size_k, int lda, int block_rows) {}
 
 template <typename scalar_t,  // compute dtype, half or nv_float16
           const vllm::ScalarTypeId w_type_id,  // weight ScalarType id
@@ -459,29 +459,32 @@ __device__ inline void barrier_release(int* lock, bool reset = false) {
 __global__ void permute_cols_kernel(int4 const* __restrict__ a_int4_ptr,
                                     int const* __restrict__ perm_int_ptr,
                                     int4* __restrict__ out_int4_ptr, int size_m,
-                                    int size_k, int block_rows) {
-  int start_row = block_rows * blockIdx.x;
+                                    int size_k, int lda, int block_rows) {
+  auto start_row = block_rows * blockIdx.x;
   int finish_row = start_row + block_rows;
   if (finish_row > size_m) {
     finish_row = size_m;
   }
   int cur_block_rows = finish_row - start_row;
 
-  int row_stride = size_k * sizeof(half) / 16;
+  int input_row_stride = lda * sizeof(half) / 16;
+  int output_row_stride = size_k * sizeof(half) / 16;
 
   auto permute_row = [&](int row) {
     int iters = size_k / default_threads;
     int rest = size_k % default_threads;
 
-    int offset = row * row_stride;
+    int input_offset = row * input_row_stride;
+    int output_offset = row * output_row_stride;
 
-    half const* a_row_half = reinterpret_cast<half const*>(a_int4_ptr + offset);
-    half* out_half = reinterpret_cast<half*>(out_int4_ptr + offset);
+    half const* a_row_half =
+        reinterpret_cast<half const*>(a_int4_ptr + input_offset);
+    half* out_half = reinterpret_cast<half*>(out_int4_ptr + output_offset);
 
     int base_k = 0;
 
     for (int i = 0; i < iters; i++) {
-      int cur_k = base_k + threadIdx.x;
+      auto cur_k = base_k + threadIdx.x;
       int src_pos = perm_int_ptr[cur_k];
 
       out_half[cur_k] = a_row_half[src_pos];
@@ -491,7 +494,7 @@ __global__ void permute_cols_kernel(int4 const* __restrict__ a_int4_ptr,
 
     if (rest) {
       if (threadIdx.x < rest) {
-        int cur_k = base_k + threadIdx.x;
+        auto cur_k = base_k + threadIdx.x;
         int src_pos = perm_int_ptr[cur_k];
 
         out_half[cur_k] = a_row_half[src_pos];
@@ -537,6 +540,7 @@ __global__ void Marlin(
     int prob_m,           // batch dimension m
     int prob_n,           // output dimension n
     int prob_k,           // reduction dimension k
+    int lda,              // A.stride(0), equal to prob_k is A is contiguous
     int* locks,           // extra global storage for barrier synchronization
     bool use_atomic_add,  // whether to use atomic add to reduce
     bool use_fp32_reduce  // whether to use fp32 global reduce
@@ -600,7 +604,7 @@ __global__ void Marlin(
   // We can easily implement parallel problem execution by just remapping
   // indices and advancing global pointers
   if (slice_col_par >= n_tiles) {
-    A += (slice_col_par / n_tiles) * 16 * thread_m_blocks * prob_k / 8;
+    A += (slice_col_par / n_tiles) * 16 * thread_m_blocks * lda / 8;
     C += (slice_col_par / n_tiles) * 16 * thread_m_blocks * prob_n / 8;
     locks += (slice_col_par / n_tiles) * n_tiles;
     slice_col = slice_col_par % n_tiles;
@@ -631,7 +635,7 @@ __global__ void Marlin(
       }
     }
     if (slice_col == n_tiles) {
-      A += 16 * thread_m_blocks * prob_k / 8;
+      A += 16 * thread_m_blocks * lda / 8;
       C += 16 * thread_m_blocks * prob_n / 8;
       locks += n_tiles;
       slice_col = 0;
@@ -643,7 +647,7 @@ __global__ void Marlin(
   // A sizes/strides
 
   // stride of the A matrix in global memory
-  int a_gl_stride = prob_k / 8;
+  int a_gl_stride = lda / 8;
   // stride of an A matrix tile in shared memory
   constexpr int a_sh_stride = 16 * thread_k_blocks / 8;
   // delta between subsequent A tiles in global memory
@@ -719,8 +723,8 @@ __global__ void Marlin(
                 (threadIdx.x % b_sh_stride_threads) * b_thread_vecs;
   b_gl_rd += b_sh_stride * slice_col;
   b_gl_rd += b_gl_rd_delta_o * slice_row;
-  int b_sh_wr = threadIdx.x * b_thread_vecs;
-  int b_sh_rd = threadIdx.x * b_thread_vecs;
+  auto b_sh_wr = threadIdx.x * b_thread_vecs;
+  auto b_sh_rd = threadIdx.x * b_thread_vecs;
 
   // For act_order
   constexpr int k_iter_size = tb_k / b_sh_wr_iters;
@@ -739,7 +743,7 @@ __global__ void Marlin(
                 s_sh_stride * slice_col + threadIdx.x;
     }
   }
-  int s_sh_wr = threadIdx.x;
+  auto s_sh_wr = threadIdx.x;
   bool s_sh_wr_pred = threadIdx.x < s_sh_stride;
 
   // Zero-points
@@ -752,7 +756,7 @@ __global__ void Marlin(
                  zp_sh_stride * slice_col + threadIdx.x;
     }
   }
-  int zp_sh_wr = threadIdx.x;
+  auto zp_sh_wr = threadIdx.x;
   bool zp_sh_wr_pred = threadIdx.x < zp_sh_stride;
 
   // We use a different scale layout for grouped and column-wise quantization as
@@ -1043,7 +1047,7 @@ __global__ void Marlin(
           int4* sh_s_stage = sh_s + s_sh_stage * pipe;
           reinterpret_cast<int4*>(&frag_s[k % 2])[0] = sh_s_stage[s_sh_rd];
         } else {
-          int warp_id = threadIdx.x / 32;
+          auto warp_id = threadIdx.x / 32;
           int n_warps = thread_n_blocks / 4;
 
           int warp_row = warp_id / n_warps;
@@ -1081,7 +1085,7 @@ __global__ void Marlin(
 
     // Determine "position" inside the thread-block (based on warp and
     // thread-id)
-    int warp_id = threadIdx.x / 32;
+    auto warp_id = threadIdx.x / 32;
     int n_warps =
         thread_n_blocks / 4;  // Each warp processes 4 16-size tiles over N
 
@@ -1090,7 +1094,7 @@ __global__ void Marlin(
 
     cur_k += warp_row * 16;
 
-    int th_id = threadIdx.x % 32;
+    auto th_id = threadIdx.x % 32;
     cur_k += (th_id % 4) * 2;  // Due to tensor-core layout for fp16 B matrix
 
     int s_col_shift =
@@ -1155,7 +1159,7 @@ __global__ void Marlin(
               (reinterpret_cast<int*>(sh_zp_stage))[zp_sh_rd + i];
         }
       } else {
-        int warp_id = threadIdx.x / 32;
+        auto warp_id = threadIdx.x / 32;
         int n_warps = thread_n_blocks / 4;
 
         int warp_row = warp_id / n_warps;
@@ -1193,7 +1197,7 @@ __global__ void Marlin(
                                      (pipe / (group_blocks / thread_k_blocks)));
           reinterpret_cast<int4*>(&frag_zpf[k % 2])[0] = sh_zp_stage[zp_sh_rd];
         } else {
-          int warp_id = threadIdx.x / 32;
+          auto warp_id = threadIdx.x / 32;
           int n_warps = thread_n_blocks / 4;
 
           int warp_row = warp_id / n_warps;
@@ -1319,7 +1323,7 @@ __global__ void Marlin(
   auto thread_block_reduce = [&]() {
     constexpr int red_off = threads / b_sh_stride_threads / 2;
     if (red_off >= 1) {
-      int red_idx = threadIdx.x / b_sh_stride_threads;
+      auto red_idx = threadIdx.x / b_sh_stride_threads;
       constexpr int red_sh_stride = b_sh_stride_threads * 4 * 2;
       constexpr int red_sh_delta = b_sh_stride_threads;
       int red_sh_rd = red_sh_stride * (threadIdx.x / b_sh_stride_threads) +
@@ -1386,7 +1390,7 @@ __global__ void Marlin(
                     4 * (threadIdx.x / 32) + threadIdx.x % 4;
       c_gl_wr += (2 * thread_n_blocks) * slice_col;
       constexpr int c_sh_wr_delta = active_threads;
-      int c_sh_wr = threadIdx.x;
+      auto c_sh_wr = threadIdx.x;
 
       int row = (threadIdx.x % 32) / 4;
 
@@ -1780,8 +1784,8 @@ __global__ void Marlin(
                HAS_ZP, GROUP_BLOCKS, IS_ZP_FLOAT>                              \
             <<<blocks, NUM_THREADS, max_shared_mem, stream>>>(                 \
                 A_ptr, B_ptr, C_ptr, C_tmp_ptr, s_ptr, zp_ptr, g_idx_ptr,      \
-                num_groups, prob_m, prob_n, prob_k, locks, use_atomic_add,     \
-                use_fp32_reduce);                                              \
+                num_groups, prob_m, prob_n, prob_k, lda, locks,                \
+                use_atomic_add, use_fp32_reduce);                              \
       }                                                                        \
     }
 
@@ -2071,7 +2075,7 @@ exec_config_t determine_thread_config(int prob_m, int prob_n, int prob_k,
 template <typename scalar_t>
 void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* s,
                void* zp, void* g_idx, void* perm, void* a_tmp, int prob_m,
-               int prob_n, int prob_k, void* workspace,
+               int prob_n, int prob_k, int lda, void* workspace,
                vllm::ScalarType const& q_type, bool has_act_order,
                bool is_k_full, bool has_zp, int num_groups, int group_size,
                int dev, cudaStream_t stream, int thread_k, int thread_n,
@@ -2184,8 +2188,9 @@ void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* s,
     // Permute A columns
     int block_rows = div_ceil(prob_m, blocks);
     permute_cols_kernel<<<blocks, default_threads, 0, stream>>>(
-        A_ptr, perm_ptr, a_tmp_ptr, prob_m, prob_k, block_rows);
+        A_ptr, perm_ptr, a_tmp_ptr, prob_m, prob_k, lda, block_rows);
     A_ptr = a_tmp_ptr;
+    lda = prob_k;
   }
 
   // If we have a full K, then we can run the non-act-order version of Marlin
@@ -2244,7 +2249,7 @@ void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* s,
                   ", num_bits = ", num_bits);
     }
 
-    A_ptr += 16 * thread_m_blocks * (prob_k / 8) * par;
+    A_ptr += 16 * thread_m_blocks * (lda / 8) * par;
     C_ptr += 16 * thread_m_blocks * (prob_n / 8) * par;
   }
 }
@@ -2300,7 +2305,10 @@ torch::Tensor gptq_marlin_gemm(torch::Tensor& a, torch::Tensor& b_q_weight,
 
   // Verify device and strides
   TORCH_CHECK(a.device().is_cuda(), "A is not on GPU");
-  TORCH_CHECK(a.is_contiguous(), "A is not contiguous");
+  TORCH_CHECK(a.stride(1) == 1, "A.stride(1) is not 1");
+  // We use int4 (16 bytes) to load A, so A must aligned to 16 bytes
+  TORCH_CHECK(a.stride(0) % 8 == 0, "A.stride(0) must divisible by 8");
+  TORCH_CHECK(((uint64_t)a.data_ptr()) % 16 == 0, "A must aligned to 16 bytes");
 
   TORCH_CHECK(b_q_weight.device().is_cuda(), "b_q_weight is not on GPU");
   TORCH_CHECK(b_q_weight.is_contiguous(), "b_q_weight is not contiguous");
@@ -2432,7 +2440,7 @@ torch::Tensor gptq_marlin_gemm(torch::Tensor& a, torch::Tensor& b_q_weight,
         a.data_ptr<at::Half>(), b_q_weight.data_ptr(), c.data_ptr<at::Half>(),
         c_tmp.data_ptr<float>(), b_scales.data_ptr<at::Half>(),
         b_zeros.data_ptr(), g_idx.data_ptr(), perm.data_ptr(),
-        a_tmp.data_ptr<at::Half>(), size_m, size_n, size_k,
+        a_tmp.data_ptr<at::Half>(), size_m, size_n, size_k, a.stride(0),
         workspace.data_ptr(), b_q_type, has_act_order, is_k_full, has_zp,
         num_groups, group_size, dev, at::cuda::getCurrentCUDAStream(dev),
         thread_k, thread_n, sms, marlin::max_par, use_atomic_add,
@@ -2443,10 +2451,10 @@ torch::Tensor gptq_marlin_gemm(torch::Tensor& a, torch::Tensor& b_q_weight,
         c.data_ptr<at::BFloat16>(), c_tmp.data_ptr<float>(),
         b_scales.data_ptr<at::BFloat16>(), b_zeros.data_ptr(), g_idx.data_ptr(),
         perm.data_ptr(), a_tmp.data_ptr<at::BFloat16>(), size_m, size_n, size_k,
-        workspace.data_ptr(), b_q_type, has_act_order, is_k_full, has_zp,
-        num_groups, group_size, dev, at::cuda::getCurrentCUDAStream(dev),
-        thread_k, thread_n, sms, marlin::max_par, use_atomic_add,
-        use_fp32_reduce, is_zp_float);
+        a.stride(0), workspace.data_ptr(), b_q_type, has_act_order, is_k_full,
+        has_zp, num_groups, group_size, dev,
+        at::cuda::getCurrentCUDAStream(dev), thread_k, thread_n, sms,
+        marlin::max_par, use_atomic_add, use_fp32_reduce, is_zp_float);
   } else {
     TORCH_CHECK(false, "gpt_marlin_gemm only supports bfloat16 and float16");
   }

csrc/quantization/gptq_marlin/gptq_marlin_repack.cu

@@ -15,7 +15,7 @@ __global__ void gptq_marlin_repack_kernel(
   int n_tiles = size_n / tile_n_size;
   int block_k_tiles = div_ceil(k_tiles, gridDim.x);
 
-  int start_k_tile = blockIdx.x * block_k_tiles;
+  auto start_k_tile = blockIdx.x * block_k_tiles;
   if (start_k_tile >= k_tiles) {
     return;
   }
@@ -71,8 +71,8 @@ __global__ void gptq_marlin_repack_kernel(
 
     if constexpr (has_perm) {
       if (threadIdx.x < stage_size) {
-        int k_id = threadIdx.x / stage_n_threads;
-        int n_id = threadIdx.x % stage_n_threads;
+        auto k_id = threadIdx.x / stage_n_threads;
+        auto n_id = threadIdx.x % stage_n_threads;
 
         uint32_t const* sh_perm_int_ptr =
             reinterpret_cast<uint32_t const*>(sh_perm_ptr);
@@ -88,8 +88,8 @@ __global__ void gptq_marlin_repack_kernel(
 
     } else {
       if (threadIdx.x < stage_size) {
-        int k_id = threadIdx.x / stage_n_threads;
-        int n_id = threadIdx.x % stage_n_threads;
+        auto k_id = threadIdx.x / stage_n_threads;
+        auto n_id = threadIdx.x % stage_n_threads;
 
         int first_k = k_tile_id * tile_k_size;
         int first_k_packed = first_k / pack_factor;
@@ -109,8 +109,8 @@ __global__ void gptq_marlin_repack_kernel(
       return;
     }
 
-    int warp_id = threadIdx.x / 32;
-    int th_id = threadIdx.x % 32;
+    auto warp_id = threadIdx.x / 32;
+    auto th_id = threadIdx.x % 32;
 
     if (warp_id >= 4) {
       return;
@@ -339,4 +339,4 @@ TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
 
 TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, Meta, m) {
   m.impl("gptq_marlin_repack", &gptq_marlin_repack_meta);
-}
+}

csrc/quantization/marlin/dense/marlin_cuda_kernel.cu

@@ -277,12 +277,12 @@ __global__ void Marlin(
       b_gl_stride * (threadIdx.x / b_sh_stride) + (threadIdx.x % b_sh_stride);
   b_gl_rd += b_sh_stride * slice_col;
   b_gl_rd += b_gl_rd_delta_o * slice_row;
-  int b_sh_wr = threadIdx.x;
-  int b_sh_rd = threadIdx.x;
+  auto b_sh_wr = threadIdx.x;
+  auto b_sh_rd = threadIdx.x;
 
   int s_gl_rd = s_gl_stride * ((thread_k_blocks * slice_row) / group_blocks) +
                 s_sh_stride * slice_col + threadIdx.x;
-  int s_sh_wr = threadIdx.x;
+  auto s_sh_wr = threadIdx.x;
   int s_sh_rd;
   // We use a different scale layout for grouped and column-wise quantization as
   // we scale a `half2` tile in column-major layout in the former and in
@@ -455,7 +455,7 @@ __global__ void Marlin(
   auto thread_block_reduce = [&]() {
     constexpr int red_off = threads / b_sh_stride / 2;
     if (red_off >= 1) {
-      int red_idx = threadIdx.x / b_sh_stride;
+      auto red_idx = threadIdx.x / b_sh_stride;
       constexpr int red_sh_stride = b_sh_stride * 4 * 2;
       constexpr int red_sh_delta = b_sh_stride;
       int red_sh_rd = red_sh_stride * (threadIdx.x / b_sh_stride) +
@@ -522,7 +522,7 @@ __global__ void Marlin(
                     4 * (threadIdx.x / 32) + threadIdx.x % 4;
       c_gl_wr += (2 * thread_n_blocks) * slice_col;
       constexpr int c_sh_wr_delta = active_threads;
-      int c_sh_wr = threadIdx.x;
+      auto c_sh_wr = threadIdx.x;
 
       int row = (threadIdx.x % 32) / 4;
 

csrc/quantization/marlin/qqq/marlin_qqq_gemm_kernel.cu

@@ -353,10 +353,10 @@ __global__ void Marlin(
       b_gl_stride * (threadIdx.x / b_sh_stride) + (threadIdx.x % b_sh_stride);
   b_gl_rd += b_sh_stride * slice_col;
   b_gl_rd += b_gl_rd_delta_o * slice_row;
-  int b_sh_wr = threadIdx.x;
-  int b_sh_rd = threadIdx.x;
+  auto b_sh_wr = threadIdx.x;
+  auto b_sh_rd = threadIdx.x;
 
-  int s_tok_gl_rd = threadIdx.x;
+  auto s_tok_gl_rd = threadIdx.x;
   // NOTE(HandH1998): activation scale s_tok need shuffle to [0, 8, 1, 9, 2, 10,
   // 3, 11, 4, 12, 5, 13, 6, 14, 7, 15] for example, 0, 8 row scales serve for
   // thread 0, 1, 2, 3. For more details, refer to mma operand A layout as
@@ -368,8 +368,8 @@ __global__ void Marlin(
   int s_tok_sh_rd = (threadIdx.x % 32) / 4;
   bool s_tok_sh_wr_pred = threadIdx.x < prob_m;
 
-  int s_ch_gl_rd = s_ch_sh_stride * slice_col + threadIdx.x;
-  int s_ch_sh_wr = threadIdx.x;
+  auto s_ch_gl_rd = s_ch_sh_stride * slice_col + threadIdx.x;
+  auto s_ch_sh_wr = threadIdx.x;
   int s_ch_sh_rd = 16 * ((threadIdx.x / 32) % (thread_n_blocks / 4)) +
                    2 * ((threadIdx.x % 32) % 4);
   bool s_ch_sh_wr_pred = threadIdx.x < s_ch_sh_stride;
@@ -558,7 +558,7 @@ __global__ void Marlin(
   auto thread_block_reduce = [&]() {
     constexpr int red_off = threads / b_sh_stride / 2;
     if (red_off >= 1) {
-      int red_idx = threadIdx.x / b_sh_stride;
+      auto red_idx = threadIdx.x / b_sh_stride;
       constexpr int red_sh_stride = b_sh_stride * 4 * 2;
       constexpr int red_sh_delta = b_sh_stride;
       int red_sh_rd = red_sh_stride * (threadIdx.x / b_sh_stride) +
@@ -628,7 +628,7 @@ __global__ void Marlin(
                     8 * (threadIdx.x / 32) + (threadIdx.x % 4) * 2;
       c_gl_wr += (4 * thread_n_blocks) * slice_col;
       constexpr int c_sh_wr_delta = active_threads * 2;
-      int c_sh_wr = 2 * threadIdx.x;
+      auto c_sh_wr = 2 * threadIdx.x;
 
       int row = (threadIdx.x % 32) / 4;
 

csrc/quantization/marlin/sparse/marlin_24_cuda_kernel.cu

@@ -273,15 +273,15 @@ __global__ void Marlin_24(
                 (threadIdx.x % b_sh_stride_threads) * b_thread_vecs;
   b_gl_rd += b_sh_stride * slice_col;
   b_gl_rd += b_gl_rd_delta_o * slice_row;
-  int b_sh_wr = threadIdx.x * b_thread_vecs;
-  int b_sh_rd = threadIdx.x * b_thread_vecs;
+  auto b_sh_wr = threadIdx.x * b_thread_vecs;
+  auto b_sh_rd = threadIdx.x * b_thread_vecs;
 
   int m_gl_rd = m_gl_stride * (threadIdx.x / (m_sh_stride)) +
                 (threadIdx.x % (m_sh_stride));
   m_gl_rd += (m_sh_stride)*slice_col;
   m_gl_rd += m_gl_rd_delta_o * slice_row;
-  int m_sh_wr = threadIdx.x;
-  int m_sh_rd = threadIdx.x % 16 + (threadIdx.x / 32) * 16;
+  auto m_sh_wr = threadIdx.x;
+  auto m_sh_rd = threadIdx.x % 16 + (threadIdx.x / 32) * 16;
 
   int s_gl_rd;
   if constexpr (group_blocks == -1) {
@@ -291,7 +291,7 @@ __global__ void Marlin_24(
               s_sh_stride * slice_col + threadIdx.x;
   }
 
-  int s_sh_wr = threadIdx.x;
+  auto s_sh_wr = threadIdx.x;
   int s_sh_rd;
   // We use a different scale layout for grouped and column-wise quantization as
   // we scale a `half2` tile in column-major layout in the former and in
@@ -516,7 +516,7 @@ __global__ void Marlin_24(
   auto thread_block_reduce = [&]() {
     constexpr int red_off = threads / b_sh_stride_threads / 2;
     if (red_off >= 1) {
-      int red_idx = threadIdx.x / b_sh_stride_threads;
+      auto red_idx = threadIdx.x / b_sh_stride_threads;
       constexpr int red_sh_stride = b_sh_stride_threads * 4 * 2;
       constexpr int red_sh_delta = b_sh_stride_threads;
       int red_sh_rd = red_sh_stride * (threadIdx.x / b_sh_stride_threads) +
@@ -583,7 +583,7 @@ __global__ void Marlin_24(
                     8 * (threadIdx.x / 32) + (threadIdx.x % 32) / 4;
       c_gl_wr += (2 * thread_n_blocks) * slice_col;
       constexpr int c_sh_wr_delta = active_threads;
-      int c_sh_wr = threadIdx.x;
+      auto c_sh_wr = threadIdx.x;
 
       int col = 2 * ((threadIdx.x % 32) % 4);
 

csrc/quantization/utils.cuh

@@ -0,0 +1,59 @@
+#pragma once
+
+/**
+ * Quantization utilities including:
+ *   Adjusted maximum values for qtypes.
+ *   Minimum scaling factors for qtypes.
+ */
+
+#include <cmath>
+#include <torch/types.h>
+
+#ifndef USE_ROCM
+  #include <c10/util/Float8_e4m3fn.h>
+  #define MAYBE_HOST_DEVICE C10_HOST_DEVICE
+#else
+  #include <ATen/hip/HIPContext.h>
+  #include <c10/util/Float8_e4m3fn.h>
+  #include <c10/util/Float8_e4m3fnuz.h>
+  // ROCm doesn't seem to need C10_HOST_DEVICE for static constexpr
+  #define MAYBE_HOST_DEVICE
+#endif
+
+template <typename T,
+          typename = std::enable_if_t<std::is_same_v<T, c10::Float8_e4m3fn> ||
+                                      std::is_same_v<T, c10::Float8_e4m3fnuz> ||
+                                      std::is_same_v<T, int8_t>>>
+struct quant_type_max {
+  static constexpr T val() { return std::numeric_limits<T>::max(); }
+};
+
+// Using the default max value from pytorch (240.0 0x7F) will cause accuracy
+// issues when running dynamic quantization. Here use 224.0 0x7E for rocm.
+template <>
+struct quant_type_max<c10::Float8_e4m3fnuz> {
+  static constexpr c10::Float8_e4m3fnuz val() {
+    return c10::Float8_e4m3fnuz(0x7E, c10::Float8_e4m3fnuz::from_bits());
+  }
+};
+
+template <typename T>
+MAYBE_HOST_DEVICE static constexpr T quant_type_max_v =
+    quant_type_max<T>::val();
+
+template <typename T,
+          typename = std::enable_if_t<std::is_same_v<T, c10::Float8_e4m3fn> ||
+                                      std::is_same_v<T, c10::Float8_e4m3fnuz> ||
+                                      std::is_same_v<T, int8_t>>>
+struct min_scaling_factor {
+  C10_DEVICE C10_ALWAYS_INLINE static float val() {
+    return 1.0f / (quant_type_max_v<T> * 512.0f);
+  }
+};
+
+template <>
+struct min_scaling_factor<int8_t> {
+  C10_DEVICE C10_ALWAYS_INLINE static float val() {
+    return std::numeric_limits<float>::epsilon();
+  }
+};

csrc/rocm/attention.cu

@@ -272,6 +272,7 @@ __launch_bounds__(NUM_THREADS, 5) void paged_attention_ll4mi_QKV_mfma16_kernel(
     const float scale,    
     const int* __restrict__ block_tables,   // [num_seqs, max_num_blocks_per_seq]
     const int* __restrict__ context_lens,   // [num_seqs]
+    const int* __restrict__ query_start_loc_ptr,   // [num_seqs]
     const int max_num_blocks_per_seq,
     const float* __restrict__ alibi_slopes, // [num_heads]
     const int q_stride,
@@ -284,18 +285,25 @@ __launch_bounds__(NUM_THREADS, 5) void paged_attention_ll4mi_QKV_mfma16_kernel(
     int max_ctx_blocks, const float* k_scale, const float* v_scale) {
   // clang-format on
   constexpr int NWARPS = NUM_THREADS / WARP_SIZE;
-  const int warpid = threadIdx.x / WARP_SIZE;
-  const int laneid = threadIdx.x % WARP_SIZE;
+  const auto warpid = threadIdx.x / WARP_SIZE;
+  const auto laneid = threadIdx.x % WARP_SIZE;
   const int lane4id = laneid % 4;
   const int lane16id = laneid % 16;
   const int rowid = laneid / 16;
 
-  const int seq_idx = blockIdx.x;
-  const int partition_idx = blockIdx.y;
+  const auto seq_idx = blockIdx.x;
+  // NOTE queries with sequence len > 1 are prefills and taken care by another
+  // kernel.
+  if (query_start_loc_ptr != nullptr &&
+      (query_start_loc_ptr[seq_idx + 1] - query_start_loc_ptr[seq_idx]) != 1) {
+    return;
+  }
+
+  const auto partition_idx = blockIdx.y;
 
   constexpr int T_PAR_SIZE = 256;  // token partition size set to 256
 
-  const int max_num_partitions = gridDim.y;
+  const auto max_num_partitions = gridDim.y;
 
   const int context_len = context_lens[seq_idx];
 
@@ -346,9 +354,9 @@ __launch_bounds__(NUM_THREADS, 5) void paged_attention_ll4mi_QKV_mfma16_kernel(
   // can be interpreted as B8x16 for 8 bit types
   _B16x8 Klocal[TLOOP][QKHELOOP];
 
-  const int wg_start_head_idx = blockIdx.z * GQA_RATIO;
-  const int wg_start_kv_head_idx = blockIdx.z;
-  const int total_num_heads = gridDim.z * GQA_RATIO;
+  const auto wg_start_head_idx = blockIdx.z * GQA_RATIO;
+  const auto wg_start_kv_head_idx = blockIdx.z;
+  const auto total_num_heads = gridDim.z * GQA_RATIO;
 
   // for QK mfma, tokens in multiples of TOKENS_PER_WARP are spread across warps
   // each mfma takes QH16xT16x16HE across warp
@@ -377,9 +385,10 @@ __launch_bounds__(NUM_THREADS, 5) void paged_attention_ll4mi_QKV_mfma16_kernel(
   // fetch Q in shared across warps and then write to registers
   const int local_qhead_idx = 4 * warpid + rowid;
   const int global_qhead_idx = wg_start_head_idx + local_qhead_idx;
-  const int64_t seq_idx64 = static_cast<int64_t>(seq_idx);
+  const int64_t query_start_off = static_cast<int64_t>(
+      query_start_loc_ptr ? query_start_loc_ptr[seq_idx] : seq_idx);
   const scalar_t* q_ptr =
-      q + seq_idx64 * q_stride + global_qhead_idx * HEAD_SIZE;
+      q + query_start_off * q_stride + global_qhead_idx * HEAD_SIZE;
 
   const int qhead_element = lane16id * CONTIGUOUS_SCALAR_ELEMS_16B;
   if ((local_qhead_idx < GQA_RATIO) && (qhead_element < HEAD_SIZE)) {
@@ -777,6 +786,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_mfma4_kernel(
     const float scale,
     const int* __restrict__ block_tables,   // [num_seqs, max_num_blocks_per_seq]
     const int* __restrict__ context_lens,   // [num_seqs]
+    const int* __restrict__ query_start_loc_ptr,   // [num_seqs]
     const int max_num_blocks_per_seq,
     const float* __restrict__ alibi_slopes, // [num_heads]
     const int q_stride,
@@ -789,14 +799,20 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_mfma4_kernel(
     int max_ctx_blocks, const float* k_scale, const float* v_scale) {
   // clang-format on
   constexpr int NWARPS = NUM_THREADS / WARP_SIZE;
-  const int warpid = threadIdx.x / WARP_SIZE;
-  const int laneid = threadIdx.x % WARP_SIZE;
+  const auto warpid = threadIdx.x / WARP_SIZE;
+  const auto laneid = threadIdx.x % WARP_SIZE;
   const int lane4id = laneid % 4;
 
-  const int seq_idx = blockIdx.x;
-  const int partition_idx = blockIdx.y;
-  const int partition_size = blockDim.x;
-  const int max_num_partitions = gridDim.y;
+  const auto seq_idx = blockIdx.x;
+  // NOTE queries with sequence len > 1 are prefills and taken care by another
+  // kernel.
+  if (query_start_loc_ptr != nullptr &&
+      (query_start_loc_ptr[seq_idx + 1] - query_start_loc_ptr[seq_idx] != 1)) {
+    return;
+  }
+  const auto partition_idx = blockIdx.y;
+  const auto partition_size = blockDim.x;
+  const auto max_num_partitions = gridDim.y;
 
   const int context_len = context_lens[seq_idx];
   const int partition_start_token_idx = partition_idx * partition_size;
@@ -838,8 +854,8 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_mfma4_kernel(
     qk_max[h] = -FLT_MAX;
   }
 
-  const int wg_start_head_idx = blockIdx.z * GQA_RATIO;
-  const int wg_start_kv_head_idx = blockIdx.z;
+  const auto wg_start_head_idx = blockIdx.z * GQA_RATIO;
+  const auto wg_start_kv_head_idx = blockIdx.z;
 
   const int warp_start_token_idx =
       partition_start_token_idx + warpid * WARP_SIZE;
@@ -857,7 +873,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_mfma4_kernel(
 
     const int* block_table = block_tables + seq_idx * max_num_blocks_per_seq;
     // token id within partition
-    const int local_token_idx = threadIdx.x;
+    const auto local_token_idx = threadIdx.x;
     // token id within sequence
     const int global_token_idx = partition_start_token_idx + local_token_idx;
 
@@ -882,9 +898,11 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_mfma4_kernel(
     }
 
     // fetch q elements
-    // every 4 lanes fetch 8 elems, so warp fetches 8*16 = 128 elems
+    // every 4 lanes fetch 8 elems, so warp fetches 8*16 = 128 elemsc
+    const int64_t query_start_off = static_cast<int64_t>(
+        query_start_loc_ptr ? query_start_loc_ptr[seq_idx] : seq_idx);
     const scalar_t* q_ptr =
-        q + seq_idx * q_stride + wg_start_head_idx * HEAD_SIZE;
+        q + query_start_off * q_stride + wg_start_head_idx * HEAD_SIZE;
     const _B16x8* q_ptrh8 = reinterpret_cast<const _B16x8*>(q_ptr);
     const int qhead_elemh8 = laneid / 4;
 
@@ -1126,7 +1144,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_mfma4_kernel(
 
   __syncthreads();
 
-  const int num_heads = gridDim.z * GQA_RATIO;
+  const auto num_heads = gridDim.z * GQA_RATIO;
   float* max_logits_ptr =
       max_logits + seq_idx * num_heads * max_num_partitions + partition_idx;
   float* exp_sums_ptr =
@@ -1267,15 +1285,24 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
     const scalar_t* __restrict__ tmp_out,  // [num_seqs, num_heads,
                                            // max_num_partitions, head_size]
     const int* __restrict__ context_lens,  // [num_seqs]
+    const int* __restrict__ query_start_loc_ptr,  // [num_seqs]
     const int max_num_partitions) {
-  const int num_heads = gridDim.x;
-  const int head_idx = blockIdx.x;
-  const int seq_idx = blockIdx.y;
+  const auto num_heads = gridDim.x;
+  const auto head_idx = blockIdx.x;
+  const auto seq_idx = blockIdx.y;
+
+  // NOTE queries with sequence len > 1 are prefills and taken care by another
+  // kernel.
+  if (query_start_loc_ptr != nullptr &&
+      (query_start_loc_ptr[seq_idx + 1] - query_start_loc_ptr[seq_idx] != 1)) {
+    return;
+  }
+
   const int context_len = context_lens[seq_idx];
   const int num_partitions = DIVIDE_ROUND_UP(context_len, PARTITION_SIZE);
   [[maybe_unused]] constexpr int NUM_WARPS = NUM_THREADS / WARP_SIZE;
-  const int warpid = threadIdx.x / WARP_SIZE;
-  [[maybe_unused]] const int laneid = threadIdx.x % WARP_SIZE;
+  const auto warpid = threadIdx.x / WARP_SIZE;
+  [[maybe_unused]] const auto laneid = threadIdx.x % WARP_SIZE;
 
   __shared__ float shared_global_exp_sum;
   // max num partitions supported is warp_size * NPAR_LOOPS
@@ -1294,7 +1321,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
 
   #pragma unroll
     for (int i = 0; i < NPAR_LOOPS; i++) {
-      const int partition_no = i * WARP_SIZE + threadIdx.x;
+      const auto partition_no = i * WARP_SIZE + threadIdx.x;
       valid_partition[i] =
           (partition_no < num_partitions) ? partition_no : last_valid_partition;
     }
@@ -1324,7 +1351,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
     }
   #pragma unroll
     for (int i = 0; i < NPAR_LOOPS; i++) {
-      const int partition_no = i * WARP_SIZE + threadIdx.x;
+      const auto partition_no = i * WARP_SIZE + threadIdx.x;
       rescaled_exp_sum[i] *= (partition_no < num_partitions)
                                  ? expf(reg_max_logit[i] - max_logit)
                                  : 0.0f;
@@ -1336,7 +1363,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
     }
   #pragma unroll
     for (int i = 0; i < NPAR_LOOPS; i++) {
-      const int partition_no = i * WARP_SIZE + threadIdx.x;
+      const auto partition_no = i * WARP_SIZE + threadIdx.x;
       shared_exp_sums[partition_no] = rescaled_exp_sum[i];
     }
 
@@ -1439,7 +1466,9 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
       __fdividef(1.0f, shared_global_exp_sum + 1e-6f);
   acc *= inv_global_exp_sum;
 
-  OUTT* out_ptr = out + static_cast<int64_t>(seq_idx) * num_heads * HEAD_SIZE +
+  const int64_t query_start_off = static_cast<int64_t>(
+      query_start_loc_ptr ? query_start_loc_ptr[seq_idx] : seq_idx);
+  OUTT* out_ptr = out + query_start_off * num_heads * HEAD_SIZE +
                   static_cast<int64_t>(head_idx) * HEAD_SIZE;
   if constexpr (std::is_same<OUTT, bit8_t>::value) {
     out_ptr[threadIdx.x] =
@@ -1466,6 +1495,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_mfma16_kernel(
     const float scale,
     const int* __restrict__ block_tables,    // [num_seqs, max_num_blocks_per_seq]
     const int* __restrict__ context_lens,    // [num_seqs]
+    const int* __restrict__ query_start_loc_ptr,  // [num_seqs]
     const int max_num_blocks_per_seq,
     const float* __restrict__ alibi_slopes,  // [num_heads]
     const int q_stride,
@@ -1492,6 +1522,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_QKV_mfma4_kernel(
     const float scale,
     const int* __restrict__ block_tables,    // [num_seqs, max_num_blocks_per_seq]
     const int* __restrict__ context_lens,    // [num_seqs]
+    const int* __restrict__ query_start_loc_ptr,  // [num_seqs]
     const int max_num_blocks_per_seq,
     const float* __restrict__ alibi_slopes,  // [num_heads]
     const int q_stride,
@@ -1515,6 +1546,7 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
     const float* __restrict__ max_logits,  // [num_seqs, num_heads, max_num_partitions]
     const scalar_t* __restrict__ tmp_out,  // [num_seqs, num_heads, max_num_partitions, head_size]
     const int* __restrict__ context_lens,  // [num_seqs]
+    const int* __restrict__ query_start_loc_ptr,  // [num_seqs]
     const int max_num_partitions) {
   UNREACHABLE_CODE
 }
@@ -1522,34 +1554,34 @@ __launch_bounds__(NUM_THREADS) void paged_attention_ll4mi_reduce_kernel(
 
 #endif  // defined(__HIP__MI300_MI250__) TODO: Add NAVI support
 
-#define LAUNCH_CUSTOM_ATTENTION_MFMA16(GQA_RATIO)                             \
-  paged_attention_ll4mi_QKV_mfma16_kernel<T, KVT, KV_DTYPE, OUTT, BLOCK_SIZE, \
-                                          HEAD_SIZE, NTHR, ALIBI_ENABLED,     \
-                                          GQA_RATIO>                          \
-      <<<grid, block, 0, stream>>>(                                           \
-          query_ptr, key_cache_ptr, value_cache_ptr, num_kv_heads, scale,     \
-          block_tables_ptr, context_lens_ptr, max_num_blocks_per_seq,         \
-          alibi_slopes_ptr, q_stride, kv_block_stride, kv_head_stride,        \
-          exp_sums_ptr, max_logits_ptr, tmp_out_ptr, out_ptr, max_ctx_blocks, \
-          k_scale_ptr, v_scale_ptr);
+#define LAUNCH_CUSTOM_ATTENTION_MFMA16(GQA_RATIO)                              \
+  paged_attention_ll4mi_QKV_mfma16_kernel<T, KVT, KV_DTYPE, OUTT, BLOCK_SIZE,  \
+                                          HEAD_SIZE, NTHR, ALIBI_ENABLED,      \
+                                          GQA_RATIO>                           \
+      <<<grid, block, 0, stream>>>(                                            \
+          query_ptr, key_cache_ptr, value_cache_ptr, num_kv_heads, scale,      \
+          block_tables_ptr, context_lens_ptr, query_start_loc_ptr,             \
+          max_num_blocks_per_seq, alibi_slopes_ptr, q_stride, kv_block_stride, \
+          kv_head_stride, exp_sums_ptr, max_logits_ptr, tmp_out_ptr, out_ptr,  \
+          max_ctx_blocks, k_scale_ptr, v_scale_ptr);
 
-#define LAUNCH_CUSTOM_ATTENTION_MFMA4(GQA_RATIO)                              \
-  paged_attention_ll4mi_QKV_mfma4_kernel<T, KVT, KV_DTYPE, OUTT, BLOCK_SIZE,  \
-                                         HEAD_SIZE, NTHR, ALIBI_ENABLED,      \
-                                         GQA_RATIO>                           \
-      <<<grid, block, 0, stream>>>(                                           \
-          query_ptr, key_cache_ptr, value_cache_ptr, num_kv_heads, scale,     \
-          block_tables_ptr, context_lens_ptr, max_num_blocks_per_seq,         \
-          alibi_slopes_ptr, q_stride, kv_block_stride, kv_head_stride,        \
-          exp_sums_ptr, max_logits_ptr, tmp_out_ptr, out_ptr, max_ctx_blocks, \
-          k_scale_ptr, v_scale_ptr);
+#define LAUNCH_CUSTOM_ATTENTION_MFMA4(GQA_RATIO)                               \
+  paged_attention_ll4mi_QKV_mfma4_kernel<T, KVT, KV_DTYPE, OUTT, BLOCK_SIZE,   \
+                                         HEAD_SIZE, NTHR, ALIBI_ENABLED,       \
+                                         GQA_RATIO>                            \
+      <<<grid, block, 0, stream>>>(                                            \
+          query_ptr, key_cache_ptr, value_cache_ptr, num_kv_heads, scale,      \
+          block_tables_ptr, context_lens_ptr, query_start_loc_ptr,             \
+          max_num_blocks_per_seq, alibi_slopes_ptr, q_stride, kv_block_stride, \
+          kv_head_stride, exp_sums_ptr, max_logits_ptr, tmp_out_ptr, out_ptr,  \
+          max_ctx_blocks, k_scale_ptr, v_scale_ptr);
 
 #define LAUNCH_CUSTOM_REDUCTION(NPAR_LOOPS)                          \
   paged_attention_ll4mi_reduce_kernel<T, OUTT, HEAD_SIZE, HEAD_SIZE, \
                                       PARTITION_SIZE, NPAR_LOOPS>    \
       <<<reduce_grid, reduce_block, 0, stream>>>(                    \
           out_ptr, exp_sums_ptr, max_logits_ptr, tmp_out_ptr,        \
-          context_lens_ptr, max_num_partitions);
+          context_lens_ptr, query_start_loc_ptr, max_num_partitions);
 
 template <typename T, typename KVT, vllm::Fp8KVCacheDataType KV_DTYPE,
           int BLOCK_SIZE, int HEAD_SIZE, typename OUTT, int PARTITION_SIZE_OLD,
@@ -1559,9 +1591,10 @@ void paged_attention_custom_launcher(
     torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
     torch::Tensor& value_cache, const int num_kv_heads, float scale,
     torch::Tensor& block_tables, torch::Tensor& context_lens,
-    int max_context_len, const std::optional<torch::Tensor>& alibi_slopes,
-    torch::Tensor& k_scale, torch::Tensor& v_scale) {
-  int num_seqs = query.size(0);
+    const std::optional<torch::Tensor>& query_start_loc, int max_context_len,
+    const std::optional<torch::Tensor>& alibi_slopes, torch::Tensor& k_scale,
+    torch::Tensor& v_scale) {
+  int num_seqs = block_tables.size(0);
   int num_heads = query.size(1);
   int head_size = query.size(2);
   int max_num_blocks_per_seq = block_tables.size(1);
@@ -1569,6 +1602,13 @@ void paged_attention_custom_launcher(
   int kv_block_stride = key_cache.stride(0);
   int kv_head_stride = key_cache.stride(1);
 
+  // NOTE: query start location is optional for V0 decode should not be used.
+  // If batch contains mix of prefills and decode, prefills should be skipped.
+  const int* query_start_loc_ptr =
+      query_start_loc
+          ? reinterpret_cast<const int*>(query_start_loc.value().data_ptr())
+          : nullptr;
+
   // NOTE: alibi_slopes is optional.
   const float* alibi_slopes_ptr =
       alibi_slopes
@@ -1700,8 +1740,8 @@ void paged_attention_custom_launcher(
   paged_attention_custom_launcher<T, KVT, KV_DTYPE, BLK_SIZE, HEAD_SIZE, T, \
                                   PSIZE, ALIBI_ENABLED>(                    \
       out, exp_sums, max_logits, tmp_out, query, key_cache, value_cache,    \
-      num_kv_heads, scale, block_tables, context_lens, max_context_len,     \
-      alibi_slopes, k_scale, v_scale);
+      num_kv_heads, scale, block_tables, context_lens, query_start_loc,     \
+      max_context_len, alibi_slopes, k_scale, v_scale);
 
 #define CALL_CUSTOM_LAUNCHER_ALIBI(T, KVT, KV_DTYPE, BLK_SIZE, HEAD_SIZE,      \
                                    PSIZE)                                      \
@@ -1750,6 +1790,7 @@ void paged_attention(
     double scale,
     torch::Tensor& block_tables, // [num_seqs, max_num_blocks_per_seq]
     torch::Tensor& context_lens, // [num_seqs]
+    const std::optional<torch::Tensor>& query_start_loc, // [num_seqs]
     int64_t block_size, int64_t max_context_len,
     const std::optional<torch::Tensor>& alibi_slopes,
     const std::string& kv_cache_dtype, torch::Tensor& k_scale,

csrc/rocm/ops.h

@@ -7,8 +7,9 @@ void paged_attention(torch::Tensor& out, torch::Tensor& exp_sums,
                      torch::Tensor& query, torch::Tensor& key_cache,
                      torch::Tensor& value_cache, int64_t num_kv_heads,
                      double scale, torch::Tensor& block_tables,
-                     torch::Tensor& context_lens, int64_t block_size,
-                     int64_t max_context_len,
+                     torch::Tensor& context_lens,
+                     const std::optional<torch::Tensor>& query_start_loc,
+                     int64_t block_size, int64_t max_context_len,
                      const std::optional<torch::Tensor>& alibi_slopes,
                      const std::string& kv_cache_dtype, torch::Tensor& k_scale,
                      torch::Tensor& v_scale);

csrc/rocm/torch_bindings.cpp

@@ -23,7 +23,9 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, rocm_ops) {
       "                Tensor query, Tensor key_cache,"
       "                Tensor value_cache, int num_kv_heads,"
       "                float scale, Tensor block_tables,"
-      "                Tensor context_lens, int block_size,"
+      "                Tensor context_lens,"
+      "                Tensor? query_start_loc,"
+      "                int block_size,"
       "                int max_context_len,"
       "                Tensor? alibi_slopes,"
       "                str kv_cache_dtype,"

csrc/torch_bindings.cpp

@@ -31,6 +31,10 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.def("weak_ref_tensor(Tensor input) -> Tensor");
   ops.impl("weak_ref_tensor", torch::kCUDA, &weak_ref_tensor);
 
+  ops.def("get_cuda_view_from_cpu_tensor(Tensor cpu_tensor) -> Tensor");
+  ops.impl("get_cuda_view_from_cpu_tensor", torch::kCPU,
+           &get_cuda_view_from_cpu_tensor);
+
   // Attention ops
   // Compute the attention between an input query and the cached
   // keys/values using PagedAttention.
@@ -291,7 +295,9 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
 #endif
 
   // Dequantization for GGML.
-  ops.def("ggml_dequantize(Tensor W, int type, SymInt m, SymInt n) -> Tensor");
+  ops.def(
+      "ggml_dequantize(Tensor W, int type, SymInt m, SymInt n, ScalarType? "
+      "dtype) -> Tensor");
   ops.impl("ggml_dequantize", torch::kCUDA, &ggml_dequantize);
 
   // mmvq kernel for GGML.
@@ -365,6 +371,35 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.def("cutlass_scaled_mm_supports_fp8(int cuda_device_capability) -> bool");
   ops.impl("cutlass_scaled_mm_supports_fp8", &cutlass_scaled_mm_supports_fp8);
 
+  // Check if cutlass grouped gemm is supported for CUDA devices of the given
+  // capability
+  ops.def("cutlass_group_gemm_supported(int cuda_device_capability) -> bool");
+  ops.impl("cutlass_group_gemm_supported", &cutlass_group_gemm_supported);
+
+  // CUTLASS w8a8 grouped GEMM
+  ops.def(
+      "cutlass_moe_mm(Tensor! out_tensors, Tensor a_tensors, Tensor b_tensors, "
+      "               Tensor a_scales, Tensor b_scales, Tensor expert_offsets, "
+      "               Tensor problem_sizes, Tensor a_strides, "
+      "               Tensor b_strides, Tensor c_strides) -> ()",
+      {stride_tag});
+  ops.impl("cutlass_moe_mm", torch::kCUDA, &cutlass_moe_mm);
+
+  // A function that computes data required to run fused MoE with w8a8 grouped
+  // GEMM. It takes topk_ids as an input, and computes expert_offsets
+  // (token start indices of each expert). In addition to this, it computes
+  // problem sizes for each expert's multiplication used by the two mms called
+  // from fused MoE operation, and arrays with permutations required to shuffle
+  // and de-shuffle the input/output of the fused operation.
+  ops.def(
+      "get_cutlass_moe_mm_data(Tensor topk_ids, Tensor! expert_offsets, "
+      "                        Tensor! problem_sizes1, Tensor! problem_sizes2, "
+      "                        Tensor! input_permutation, "
+      "                        Tensor! output_permutation, int num_experts, "
+      "                        int n, int k) -> ()",
+      {stride_tag});
+  ops.impl("get_cutlass_moe_mm_data", torch::kCUDA, &get_cutlass_moe_mm_data);
+
   // Check if cutlass scaled_mm supports block quantization (used by DeepSeekV3)
   ops.def(
       "cutlass_scaled_mm_supports_block_fp8(int cuda_device_capability) -> "
@@ -581,12 +616,11 @@ TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cuda_utils), cuda_utils) {
                   &get_max_shared_memory_per_block_device_attribute);
 }
 
-#ifndef USE_ROCM
 TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _custom_ar), custom_ar) {
   // Custom all-reduce kernels
   custom_ar.def(
       "init_custom_ar(int[] ipc_tensors, Tensor rank_data, "
-      "int rank, bool full_nvlink) -> int");
+      "int rank, bool fully_connected) -> int");
   custom_ar.impl("init_custom_ar", torch::kCUDA, &init_custom_ar);
   custom_ar.def(
       "all_reduce(int fa, Tensor inp, Tensor! out, int reg_buffer, "
@@ -599,7 +633,13 @@ TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _custom_ar), custom_ar) {
   custom_ar.def("register_buffer", &register_buffer);
   custom_ar.def("get_graph_buffer_ipc_meta", &get_graph_buffer_ipc_meta);
   custom_ar.def("register_graph_buffers", &register_graph_buffers);
+
+  custom_ar.def("allocate_shared_buffer_and_handle",
+                &allocate_shared_buffer_and_handle);
+  custom_ar.def("open_mem_handle(Tensor mem_handle) -> int", &open_mem_handle);
+  custom_ar.impl("open_mem_handle", torch::kCPU, &open_mem_handle);
+
+  custom_ar.def("free_shared_buffer", &free_shared_buffer);
 }
-#endif
 
 REGISTER_EXTENSION(TORCH_EXTENSION_NAME)

docker/Dockerfile

@@ -14,17 +14,22 @@ ARG PYTHON_VERSION=3.12
 ARG TARGETPLATFORM
 ENV DEBIAN_FRONTEND=noninteractive
 
-# Install minimal dependencies and uv
-RUN apt-get update -y \
-    && apt-get install -y ccache git curl wget sudo \
-    && curl -LsSf https://astral.sh/uv/install.sh | sh
-
-# Add uv to PATH
-ENV PATH="/root/.local/bin:$PATH"
-# Create venv with specified Python and activate by placing at the front of path
-ENV VIRTUAL_ENV="/opt/venv"
-RUN uv venv --python ${PYTHON_VERSION} --seed ${VIRTUAL_ENV}
-ENV PATH="$VIRTUAL_ENV/bin:$PATH"
+# Install Python and other dependencies
+RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
+    && echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections \
+    && apt-get update -y \
+    && apt-get install -y ccache software-properties-common git curl sudo \
+    && add-apt-repository ppa:deadsnakes/ppa \
+    && apt-get update -y \
+    && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
+    && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
+    && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
+    && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
+    && curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION} \
+    && python3 --version && python3 -m pip --version
+# Install uv for faster pip installs
+RUN --mount=type=cache,target=/root/.cache/uv \
+    python3 -m pip install uv
 
 # This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
 # Reference: https://github.com/astral-sh/uv/pull/1694
@@ -46,19 +51,22 @@ RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
 
 WORKDIR /workspace
 
+# install build and runtime dependencies
+
 # arm64 (GH200) build follows the practice of "use existing pytorch" build,
 # we need to install torch and torchvision from the nightly builds first,
 # pytorch will not appear as a vLLM dependency in all of the following steps
 # after this step
 RUN --mount=type=cache,target=/root/.cache/uv \
     if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
-        uv pip install --index-url https://download.pytorch.org/whl/nightly/cu126 "torch==2.7.0.dev20250121+cu126" "torchvision==0.22.0.dev20250121";  \
+        uv pip install --system --index-url https://download.pytorch.org/whl/nightly/cu128 "torch==2.8.0.dev20250318+cu128" "torchvision==0.22.0.dev20250319";  \
+        uv pip install --system --index-url https://download.pytorch.org/whl/nightly/cu128 --pre pytorch_triton==3.3.0+gitab727c40; \
     fi
 
 COPY requirements/common.txt requirements/common.txt
 COPY requirements/cuda.txt requirements/cuda.txt
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install -r requirements/cuda.txt
+    uv pip install --system -r requirements/cuda.txt
 
 # cuda arch list used by torch
 # can be useful for both `dev` and `test`
@@ -83,7 +91,7 @@ COPY requirements/build.txt requirements/build.txt
 ENV UV_HTTP_TIMEOUT=500
 
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install -r requirements/build.txt
+    uv pip install --system -r requirements/build.txt
 
 COPY . .
 ARG GIT_REPO_CHECK=0
@@ -155,7 +163,7 @@ COPY requirements/lint.txt requirements/lint.txt
 COPY requirements/test.txt requirements/test.txt
 COPY requirements/dev.txt requirements/dev.txt
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install -r requirements/dev.txt
+    uv pip install --system -r requirements/dev.txt
 #################### DEV IMAGE ####################
 
 #################### vLLM installation IMAGE ####################
@@ -171,18 +179,23 @@ ARG TARGETPLATFORM
 RUN PYTHON_VERSION_STR=$(echo ${PYTHON_VERSION} | sed 's/\.//g') && \
     echo "export PYTHON_VERSION_STR=${PYTHON_VERSION_STR}" >> /etc/environment
 
-# Install minimal dependencies and uv
-RUN apt-get update -y \
-    && apt-get install -y ccache git curl wget sudo vim \
-    && apt-get install -y ffmpeg libsm6 libxext6 libgl1 libibverbs-dev \
-    && curl -LsSf https://astral.sh/uv/install.sh | sh
-
-# Add uv to PATH
-ENV PATH="/root/.local/bin:$PATH"
-# Create venv with specified Python and activate by placing at the front of path
-ENV VIRTUAL_ENV="/opt/venv"
-RUN uv venv --python ${PYTHON_VERSION} --seed ${VIRTUAL_ENV}
-ENV PATH="$VIRTUAL_ENV/bin:$PATH"
+# Install Python and other dependencies
+RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
+    && echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections \
+    && apt-get update -y \
+    && apt-get install -y ccache software-properties-common git curl wget sudo vim python3-pip \
+    && apt-get install -y ffmpeg libsm6 libxext6 libgl1 \
+    && add-apt-repository ppa:deadsnakes/ppa \
+    && apt-get update -y \
+    && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv libibverbs-dev \
+    && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
+    && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
+    && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
+    && curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION} \
+    && python3 --version && python3 -m pip --version
+# Install uv for faster pip installs
+RUN --mount=type=cache,target=/root/.cache/uv \
+    python3 -m pip install uv
 
 # This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
 # Reference: https://github.com/astral-sh/uv/pull/1694
@@ -200,13 +213,14 @@ RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
 # after this step
 RUN --mount=type=cache,target=/root/.cache/uv \
     if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
-        uv pip install --index-url https://download.pytorch.org/whl/nightly/cu124 "torch==2.6.0.dev20241210+cu124" "torchvision==0.22.0.dev20241215";  \
+        uv pip install --system --index-url https://download.pytorch.org/whl/nightly/cu128 "torch==2.8.0.dev20250318+cu128" "torchvision==0.22.0.dev20250319";  \
+        uv pip install --system --index-url https://download.pytorch.org/whl/nightly/cu128 --pre pytorch_triton==3.3.0+gitab727c40; \
     fi
 
 # Install vllm wheel first, so that torch etc will be installed.
 RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist \
     --mount=type=cache,target=/root/.cache/uv \
-    uv pip install dist/*.whl --verbose
+    uv pip install --system dist/*.whl --verbose
 
 # If we need to build FlashInfer wheel before its release:
 # $ export FLASHINFER_ENABLE_AOT=1
@@ -221,8 +235,9 @@ RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist
 # $ # upload the wheel to a public location, e.g. https://wheels.vllm.ai/flashinfer/524304395bd1d8cd7d07db083859523fcaa246a4/flashinfer_python-0.2.1.post1+cu124torch2.5-cp38-abi3-linux_x86_64.whl
 
 RUN --mount=type=cache,target=/root/.cache/uv \
+. /etc/environment && \
 if [ "$TARGETPLATFORM" != "linux/arm64" ]; then \
-    uv pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.2.1.post2/flashinfer_python-0.2.1.post2+cu124torch2.6-cp38-abi3-linux_x86_64.whl ; \
+    uv pip install --system https://github.com/flashinfer-ai/flashinfer/releases/download/v0.2.1.post2/flashinfer_python-0.2.1.post2+cu124torch2.6-cp38-abi3-linux_x86_64.whl ; \
 fi
 COPY examples examples
 
@@ -232,7 +247,7 @@ COPY examples examples
 # TODO: Remove this once FlashInfer AOT wheel is fixed
 COPY requirements/build.txt requirements/build.txt
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install -r requirements/build.txt
+    uv pip install --system -r requirements/build.txt
 
 #################### vLLM installation IMAGE ####################
 
@@ -249,15 +264,15 @@ ENV UV_HTTP_TIMEOUT=500
 
 # install development dependencies (for testing)
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install -r requirements/dev.txt
+    uv pip install --system -r requirements/dev.txt
 
 # install development dependencies (for testing)
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install -e tests/vllm_test_utils
+    uv pip install --system -e tests/vllm_test_utils
 
 # enable fast downloads from hf (for testing)
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install hf_transfer
+    uv pip install --system hf_transfer
 ENV HF_HUB_ENABLE_HF_TRANSFER 1
 
 # Copy in the v1 package for testing (it isn't distributed yet)
@@ -282,9 +297,9 @@ ENV UV_HTTP_TIMEOUT=500
 # install additional dependencies for openai api server
 RUN --mount=type=cache,target=/root/.cache/uv \
     if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
-        uv pip install accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.42.0' 'timm==0.9.10' boto3 runai-model-streamer runai-model-streamer[s3]; \
+        uv pip install --system accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.42.0' 'timm==0.9.10' boto3 runai-model-streamer runai-model-streamer[s3]; \
     else \
-        uv pip install accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.45.0' 'timm==0.9.10' boto3 runai-model-streamer runai-model-streamer[s3]; \
+        uv pip install --system accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.45.3' 'timm==0.9.10' boto3 runai-model-streamer runai-model-streamer[s3]; \
     fi
 
 ENV VLLM_USAGE_SOURCE production-docker-image

docker/Dockerfile.cpu

@@ -0,0 +1,138 @@
+# This vLLM Dockerfile is used to construct image that can build and run vLLM on x86 CPU platform.
+#
+# Build targets:
+#   vllm-openai (default): used for serving deployment
+#   vllm-test: used for CI tests
+#   vllm-dev: used for development
+#
+# Build arguments:
+#   PYTHON_VERSION=3.12 (default)|3.11|3.10|3.9
+#   VLLM_CPU_DISABLE_AVX512=false (default)|true
+#
+
+######################### BASE IMAGE #########################
+FROM ubuntu:22.04 AS base
+
+WORKDIR /workspace/
+
+ARG PYTHON_VERSION=3.12
+ARG PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu"
+
+# Install minimal dependencies and uv
+RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
+    --mount=type=cache,target=/var/lib/apt,sharing=locked \
+    apt-get update -y \
+    && apt-get install -y --no-install-recommends ccache git curl wget ca-certificates \
+        gcc-12 g++-12 libtcmalloc-minimal4 libnuma-dev ffmpeg libsm6 libxext6 libgl1 \
+    && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 10 --slave /usr/bin/g++ g++ /usr/bin/g++-12 \
+    && curl -LsSf https://astral.sh/uv/install.sh | sh
+
+ENV CCACHE_DIR=/root/.cache/ccache
+ENV CMAKE_CXX_COMPILER_LAUNCHER=ccache
+
+ENV PATH="/root/.local/bin:$PATH"
+ENV VIRTUAL_ENV="/opt/venv"
+RUN uv venv --python ${PYTHON_VERSION} --seed ${VIRTUAL_ENV}
+ENV PATH="$VIRTUAL_ENV/bin:$PATH"
+
+ENV UV_HTTP_TIMEOUT=500
+
+# Install Python dependencies 
+ENV PIP_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
+ENV UV_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
+ENV UV_INDEX_STRATEGY="unsafe-best-match"
+ENV UV_LINK_MODE="copy"
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=bind,src=requirements/common.txt,target=requirements/common.txt \
+    --mount=type=bind,src=requirements/cpu.txt,target=requirements/cpu.txt \
+    uv pip install --upgrade pip && \
+    uv pip install -r requirements/cpu.txt
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install intel-openmp==2024.2.1 intel_extension_for_pytorch==2.6.0
+
+ENV LD_PRELOAD="/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:/opt/venv/lib/libiomp5.so:$LD_PRELOAD"
+
+RUN echo 'ulimit -c 0' >> ~/.bashrc
+
+######################### BUILD IMAGE #########################
+FROM base AS vllm-build
+
+ARG GIT_REPO_CHECK=0
+# Support for building with non-AVX512 vLLM: docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" ...
+ARG VLLM_CPU_DISABLE_AVX512
+ENV VLLM_CPU_DISABLE_AVX512=${VLLM_CPU_DISABLE_AVX512}
+
+WORKDIR /workspace/vllm
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=bind,src=requirements/build.txt,target=requirements/build.txt \
+    uv pip install -r requirements/build.txt
+
+COPY . .
+RUN --mount=type=bind,source=.git,target=.git \
+    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh ; fi
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=cache,target=/root/.cache/ccache \
+    --mount=type=bind,source=.git,target=.git \
+    VLLM_TARGET_DEVICE=cpu python3 setup.py bdist_wheel 
+
+######################### DEV IMAGE #########################
+FROM vllm-build AS vllm-dev
+
+WORKDIR /workspace/vllm
+
+RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
+    --mount=type=cache,target=/var/lib/apt,sharing=locked \
+    apt-get install -y --no-install-recommends vim numactl
+
+# install development dependencies (for testing)
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install -e tests/vllm_test_utils 
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=cache,target=/root/.cache/ccache \
+    --mount=type=bind,source=.git,target=.git \
+    VLLM_TARGET_DEVICE=cpu python3 setup.py develop 
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install -r requirements/dev.txt && \
+    pre-commit install --hook-type pre-commit --hook-type commit-msg
+
+ENTRYPOINT ["bash"]
+
+######################### TEST IMAGE #########################
+FROM base AS vllm-test
+
+WORKDIR /workspace/
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=bind,src=requirements/test.txt,target=requirements/test.txt \
+    uv pip install -r requirements/test.txt
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=bind,from=vllm-build,src=/workspace/vllm/dist,target=dist \
+    uv pip install dist/*.whl
+
+ADD ./tests/ ./tests/
+ADD ./examples/ ./examples/
+ADD ./benchmarks/ ./benchmarks/
+
+# install development dependencies (for testing)
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install -e tests/vllm_test_utils 
+
+ENTRYPOINT ["bash"]
+
+######################### RELEASE IMAGE #########################
+FROM base AS vllm-openai
+
+WORKDIR /workspace/
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=cache,target=/root/.cache/ccache \
+    --mount=type=bind,from=vllm-build,src=/workspace/vllm/dist,target=dist \
+    uv pip install dist/*.whl
+
+ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]

docker/Dockerfile.ppc64le

@@ -0,0 +1,267 @@
+ARG BASE_UBI_IMAGE_TAG=9.5-1741850109
+
+###############################################################
+# base stage with basic dependencies
+###############################################################
+
+FROM registry.access.redhat.com/ubi9/ubi-minimal:${BASE_UBI_IMAGE_TAG} AS base-builder
+
+ARG PYTHON_VERSION=3.12
+ARG OPENBLAS_VERSION=0.3.29
+
+# Set Environment Variables for venv, cargo & openblas
+ENV VIRTUAL_ENV=/opt/vllm
+ENV PATH=${VIRTUAL_ENV}/bin:/root/.cargo/bin:$PATH
+ENV PKG_CONFIG_PATH=/usr/local/lib/pkgconfig/
+ENV LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/lib64:/usr/local/lib:/usr/lib64:/usr/lib
+ENV UV_LINK_MODE=copy
+
+# install gcc-13, python, rust, openblas
+# Note: A symlink for libatomic.so is created for gcc-13 (linker fails to find libatomic otherwise - reqd. for sentencepiece)
+# Note: A dummy file 'control' is created in /tmp/ to artificially create dependencies between stages when building stages in parallel
+#       when `--jobs=<N>` is passed with podman build command
+RUN microdnf install -y openssl-devel dnf \
+    && dnf install -y https://mirror.stream.centos.org/9-stream/BaseOS/`arch`/os/Packages/centos-gpg-keys-9.0-24.el9.noarch.rpm \
+        https://mirror.stream.centos.org/9-stream/BaseOS/`arch`/os/Packages/centos-stream-repos-9.0-24.el9.noarch.rpm \
+        https://dl.fedoraproject.org/pub/epel/epel-release-latest-9.noarch.rpm \
+    && dnf config-manager --add-repo https://mirror.stream.centos.org/9-stream/BaseOS/`arch`/os \
+    && dnf config-manager --add-repo https://mirror.stream.centos.org/9-stream/AppStream/`arch`/os \
+    && dnf config-manager --set-enabled crb \
+    && dnf install -y \
+       git tar gcc-toolset-13 automake libtool numactl-devel lapack-devel \
+       pkgconfig xsimd zeromq-devel kmod findutils protobuf* \
+       libtiff-devel libjpeg-devel openjpeg2-devel zlib-devel \
+       freetype-devel lcms2-devel libwebp-devel tcl-devel tk-devel \
+       harfbuzz-devel fribidi-devel libraqm-devel libimagequant-devel libxcb-devel \
+       python${PYTHON_VERSION}-devel python${PYTHON_VERSION}-pip \
+    && dnf clean all \
+    && ln -sf /usr/lib64/libatomic.so.1 /usr/lib64/libatomic.so \
+    && python${PYTHON_VERSION} -m venv ${VIRTUAL_ENV} \
+    && python -m pip install -U pip uv \
+    && uv pip install wheel build "setuptools<70" setuptools_scm setuptools_rust meson-python 'cmake<4' ninja cython scikit_build_core scikit_build \
+    && curl -sL https://ftp2.osuosl.org/pub/ppc64el/openblas/latest/Openblas_${OPENBLAS_VERSION}_ppc64le.tar.gz | tar xvf - -C /usr/local \
+    && curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y \
+    && cd /tmp && touch control
+
+###############################################################
+# Stage to build torch family
+###############################################################
+
+FROM base-builder AS torch-builder
+
+ARG MAX_JOBS
+ARG TORCH_VERSION=2.6.0
+ARG _GLIBCXX_USE_CXX11_ABI=1
+RUN --mount=type=cache,target=/root/.cache/uv \
+    source /opt/rh/gcc-toolset-13/enable &&  \
+    git clone --recursive https://github.com/pytorch/pytorch.git -b v${TORCH_VERSION} && \
+    cd pytorch && \
+    uv pip install -r requirements.txt && \
+    python setup.py develop && \
+    rm -f dist/torch*+git*whl && \
+    MAX_JOBS=${MAX_JOBS:-$(nproc)} \
+    PYTORCH_BUILD_VERSION=${TORCH_VERSION} PYTORCH_BUILD_NUMBER=1 uv build --wheel --out-dir /torchwheels/
+
+ARG TORCHVISION_VERSION=0.21.0
+ARG TORCHVISION_USE_NVJPEG=0
+ARG TORCHVISION_USE_FFMPEG=0
+RUN --mount=type=cache,target=/root/.cache/uv \
+    source /opt/rh/gcc-toolset-13/enable && \
+    git clone --recursive https://github.com/pytorch/vision.git -b v${TORCHVISION_VERSION} && \
+    cd vision && \
+    MAX_JOBS=${MAX_JOBS:-$(nproc)} \
+    BUILD_VERSION=${TORCHVISION_VERSION} \
+    uv build --wheel --out-dir /torchwheels/ --no-build-isolation
+
+ARG TORCHAUDIO_VERSION=2.6.0
+ARG BUILD_SOX=1
+ARG BUILD_KALDI=1
+ARG BUILD_RNNT=1
+ARG USE_FFMPEG=0
+ARG USE_ROCM=0
+ARG USE_CUDA=0
+ARG TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_FFMPEG=1
+RUN --mount=type=cache,target=/root/.cache/uv \
+    source /opt/rh/gcc-toolset-13/enable && \
+    git clone --recursive https://github.com/pytorch/audio.git -b v${TORCHAUDIO_VERSION} && \
+    cd audio && \
+    MAX_JOBS=${MAX_JOBS:-$(nproc)} \
+    BUILD_VERSION=${TORCHAUDIO_VERSION} \
+    uv build --wheel --out-dir /torchwheels/ --no-build-isolation
+
+###############################################################
+# Stage to build pyarrow
+###############################################################
+
+FROM base-builder AS arrow-builder
+
+ARG MAX_JOBS
+ARG PYARROW_PARALLEL
+ARG PYARROW_VERSION=19.0.1
+RUN --mount=type=cache,target=/root/.cache/uv \
+    source /opt/rh/gcc-toolset-13/enable && \
+    git clone --recursive https://github.com/apache/arrow.git -b apache-arrow-${PYARROW_VERSION} && \
+    cd arrow/cpp && \
+    mkdir build && cd build && \
+    cmake -DCMAKE_BUILD_TYPE=release \
+        -DCMAKE_INSTALL_PREFIX=/usr/local \
+        -DARROW_PYTHON=ON \
+        -DARROW_BUILD_TESTS=OFF \
+        -DARROW_JEMALLOC=ON \
+        -DARROW_BUILD_STATIC="OFF" \
+        -DARROW_PARQUET=ON \
+        .. && \
+    make install -j ${MAX_JOBS:-$(nproc)} && \
+    cd ../../python/ && \
+    uv pip install -v -r requirements-wheel-build.txt && \
+    PYARROW_PARALLEL=${PYARROW_PARALLEL:-$(nproc)} \
+    python setup.py build_ext \
+    --build-type=release --bundle-arrow-cpp \
+    bdist_wheel --dist-dir /arrowwheels/
+
+###############################################################
+# Stage to build opencv
+###############################################################
+
+FROM base-builder AS cv-builder
+
+ARG MAX_JOBS
+ARG OPENCV_VERSION=84
+ARG ENABLE_HEADLESS=1
+RUN --mount=type=cache,target=/root/.cache/uv \
+    source /opt/rh/gcc-toolset-13/enable && \
+    git clone --recursive https://github.com/opencv/opencv-python.git -b ${OPENCV_VERSION} && \
+    cd opencv-python && \
+    sed -i 's/"setuptools==59.2.0",/"setuptools<70.0",/g' pyproject.toml && \
+    python -m build --wheel --installer=uv --outdir /opencvwheels/
+
+###############################################################
+# Stage to build vllm - this stage builds and installs
+# vllm, tensorizer and vllm-tgis-adapter and builds uv cache
+# for transitive dependencies - eg. grpcio
+###############################################################
+
+FROM base-builder AS vllmcache-builder
+
+COPY --from=torch-builder /tmp/control /dev/null
+COPY --from=arrow-builder /tmp/control /dev/null
+COPY --from=cv-builder /tmp/control /dev/null
+
+ARG VLLM_TARGET_DEVICE=cpu
+
+# this step installs vllm and populates uv cache
+# with all the transitive dependencies
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=bind,from=torch-builder,source=/torchwheels/,target=/torchwheels/,ro \
+    --mount=type=bind,from=arrow-builder,source=/arrowwheels/,target=/arrowwheels/,ro \
+    --mount=type=bind,from=cv-builder,source=/opencvwheels/,target=/opencvwheels/,ro \
+    --mount=type=bind,src=.,dst=/src/,rw \
+    source /opt/rh/gcc-toolset-13/enable && \
+    uv pip install /opencvwheels/*.whl /arrowwheels/*.whl /torchwheels/*.whl && \
+    sed -i -e 's/.*torch.*//g' /src/pyproject.toml /src/requirements/*.txt && \
+    uv pip install pandas pythran pybind11 && \
+    # sentencepiece.pc is in some pkgconfig inside uv cache
+    export PKG_CONFIG_PATH=$(find / -type d -name "pkgconfig" 2>/dev/null | tr '\n' ':') && \
+    uv pip install -r /src/requirements/common.txt -r /src/requirements/cpu.txt -r /src/requirements/build.txt --no-build-isolation && \
+    cd /src/ && \
+    uv build --wheel --out-dir /vllmwheel/ --no-build-isolation && \
+    uv pip install /vllmwheel/*.whl
+
+
+###############################################################
+# Stage to build numactl
+###############################################################
+
+FROM base-builder AS numa-builder
+
+# Note: Building numactl with gcc-11. Compiling with gcc-13 in this builder stage will
+# trigger recompilation with gcc-11 (and require libtool) in the final stage where we do not have gcc-13
+ARG MAX_JOBS
+ARG NUMACTL_VERSION=2.0.19
+RUN git clone --recursive https://github.com/numactl/numactl.git -b v${NUMACTL_VERSION} \
+    && cd numactl \
+    && autoreconf -i && ./configure \
+    && make -j ${MAX_JOBS:-$(nproc)}
+
+###############################################################
+# Stage to build lapack
+###############################################################
+
+FROM base-builder AS lapack-builder
+
+ARG MAX_JOBS
+ARG LAPACK_VERSION=3.12.1
+RUN git clone --recursive https://github.com/Reference-LAPACK/lapack.git -b v${LAPACK_VERSION} \
+    && cd lapack && source /opt/rh/gcc-toolset-13/enable \
+    && cmake -B build -S . \
+    && cmake --build build -j ${MAX_JOBS:-$(nproc)}
+
+
+###############################################################
+#                   FINAL VLLM IMAGE STAGE                    #
+###############################################################
+
+FROM registry.access.redhat.com/ubi9/ubi-minimal:${BASE_UBI_IMAGE_TAG} AS vllm-openai
+
+ARG PYTHON_VERSION=3.12
+ARG OPENBLAS_VERSION=0.3.29
+
+# Set Environment Variables for venv & openblas
+ENV VIRTUAL_ENV=/opt/vllm
+ENV PATH=${VIRTUAL_ENV}/bin:$PATH
+ENV PKG_CONFIG_PATH=/usr/local/lib/pkgconfig/
+ENV LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/lib64:/usr/local/lib:/usr/lib64:/usr/lib
+ENV UV_LINK_MODE=copy
+
+# create artificial dependencies between stages for independent stages to build in parallel
+COPY --from=torch-builder /tmp/control /dev/null
+COPY --from=arrow-builder /tmp/control /dev/null
+COPY --from=cv-builder /tmp/control /dev/null
+COPY --from=vllmcache-builder /tmp/control /dev/null
+COPY --from=numa-builder /tmp/control /dev/null
+COPY --from=lapack-builder /tmp/control /dev/null
+
+# install gcc-11, python, openblas, numactl, lapack
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=bind,from=numa-builder,source=/numactl/,target=/numactl/,rw \
+    --mount=type=bind,from=lapack-builder,source=/lapack/,target=/lapack/,rw \
+    rpm -ivh https://dl.fedoraproject.org/pub/epel/epel-release-latest-9.noarch.rpm && \
+    microdnf install --nodocs -y \
+    tar findutils openssl \
+    pkgconfig xsimd g++ gcc-fortran libsndfile \
+    libtiff libjpeg openjpeg2 zlib zeromq \
+    freetype lcms2 libwebp tcl tk utf8proc \
+    harfbuzz fribidi libraqm libimagequant libxcb \
+    python${PYTHON_VERSION}-devel python${PYTHON_VERSION}-pip \
+    && microdnf clean all \
+    && python${PYTHON_VERSION} -m venv ${VIRTUAL_ENV} \
+    && python -m pip install -U pip uv --no-cache \
+    && curl -sL https://ftp2.osuosl.org/pub/ppc64el/openblas/latest/Openblas_${OPENBLAS_VERSION}_ppc64le.tar.gz | tar xvf - -C /usr/local \
+    && make -C /numactl install \
+    && uv pip install 'cmake<4' \
+    && cmake --install /lapack/build \
+    && uv pip uninstall cmake
+
+# consume previously built wheels (including vllm)
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=bind,from=torch-builder,source=/torchwheels/,target=/torchwheels/,ro \
+    --mount=type=bind,from=arrow-builder,source=/arrowwheels/,target=/arrowwheels/,ro \
+    --mount=type=bind,from=cv-builder,source=/opencvwheels/,target=/opencvwheels/,ro \
+    --mount=type=bind,from=vllmcache-builder,source=/vllmwheel/,target=/vllmwheel/,ro \
+    HOME=/root uv pip install /opencvwheels/*.whl /arrowwheels/*.whl /torchwheels/*.whl /vllmwheel/*.whl
+
+COPY ./ /workspace/vllm
+WORKDIR /workspace/vllm
+ARG GIT_REPO_CHECK=0
+RUN --mount=type=bind,source=.git,target=.git \
+    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh; fi
+
+# install development dependencies (for testing)
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install -e tests/vllm_test_utils
+
+WORKDIR /workspace/
+
+RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
+
+ENTRYPOINT ["python", "-m", "vllm.entrypoints.openai.api_server"]

docker/Dockerfile.rocm

@@ -12,7 +12,8 @@ ENV PYTORCH_ROCM_ARCH=${ARG_PYTORCH_ROCM_ARCH:-${PYTORCH_ROCM_ARCH}}
 
 # Install some basic utilities
 RUN apt-get update -q -y && apt-get install -q -y \
-    sqlite3 libsqlite3-dev libfmt-dev libmsgpack-dev libsuitesparse-dev
+    sqlite3 libsqlite3-dev libfmt-dev libmsgpack-dev libsuitesparse-dev \
+    apt-transport-https ca-certificates wget curl
 # Remove sccache    
 RUN python3 -m pip install --upgrade pip && pip install setuptools_scm
 RUN apt-get purge -y sccache; python3 -m pip uninstall -y sccache; rm -f "$(which sccache)"
@@ -40,7 +41,7 @@ ARG USE_CYTHON
 RUN cd vllm \
     && python3 -m pip install -r requirements/rocm.txt \
     && python3 setup.py clean --all  \
-    && if [ ${USE_CYTHON} -eq "1" ]; then python3 setup_cython.py build_ext --inplace; fi \
+    && if [ ${USE_CYTHON} -eq "1" ]; then python3 tests/build_cython.py build_ext --inplace; fi \
     && python3 setup.py bdist_wheel --dist-dir=dist
 FROM scratch AS export_vllm
 ARG COMMON_WORKDIR

docker/Dockerfile.rocm_base

@@ -1,24 +1,26 @@
 ARG BASE_IMAGE=rocm/dev-ubuntu-22.04:6.3.1-complete
-ARG HIPBLASLT_BRANCH="4d40e36"
+ARG HIPBLASLT_BRANCH="db8e93b4"
 ARG HIPBLAS_COMMON_BRANCH="7c1566b"
 ARG LEGACY_HIPBLASLT_OPTION=
 ARG RCCL_BRANCH="648a58d"
 ARG RCCL_REPO="https://github.com/ROCm/rccl"
 ARG TRITON_BRANCH="e5be006"
 ARG TRITON_REPO="https://github.com/triton-lang/triton.git"
-ARG PYTORCH_BRANCH="3a585126"
-ARG PYTORCH_VISION_BRANCH="v0.19.1"
+ARG PYTORCH_BRANCH="295f2ed4"
+ARG PYTORCH_VISION_BRANCH="v0.21.0"
 ARG PYTORCH_REPO="https://github.com/pytorch/pytorch.git"
 ARG PYTORCH_VISION_REPO="https://github.com/pytorch/vision.git"
-ARG FA_BRANCH="b7d29fb"
-ARG FA_REPO="https://github.com/ROCm/flash-attention.git"
+ARG FA_BRANCH="1a7f4dfa"
+ARG FA_REPO="https://github.com/Dao-AILab/flash-attention.git"
+ARG AITER_BRANCH="8970b25b"
+ARG AITER_REPO="https://github.com/ROCm/aiter.git"
 
 FROM ${BASE_IMAGE} AS base
 
 ENV PATH=/opt/rocm/llvm/bin:$PATH
 ENV ROCM_PATH=/opt/rocm
 ENV LD_LIBRARY_PATH=/opt/rocm/lib:/usr/local/lib:
-ARG PYTORCH_ROCM_ARCH=gfx90a;gfx942
+ARG PYTORCH_ROCM_ARCH=gfx90a;gfx942;gfx1100;gfx1101;gfx1200;gfx1201
 ENV PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH}
 
 ARG PYTHON_VERSION=3.12
@@ -29,7 +31,7 @@ ENV DEBIAN_FRONTEND=noninteractive
 
 # Install Python and other dependencies
 RUN apt-get update -y \
-    && apt-get install -y software-properties-common git curl sudo vim less \
+    && apt-get install -y software-properties-common git curl sudo vim less libgfortran5 \
     && add-apt-repository ppa:deadsnakes/ppa \
     && apt-get update -y \
     && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
@@ -40,7 +42,7 @@ RUN apt-get update -y \
     && curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION} \
     && python3 --version && python3 -m pip --version
 
-RUN pip install -U packaging cmake ninja wheel setuptools pybind11 Cython
+RUN pip install -U packaging 'cmake<4' ninja wheel setuptools pybind11 Cython
 
 FROM base AS build_hipblaslt
 ARG HIPBLASLT_BRANCH
@@ -58,7 +60,8 @@ RUN cd hipBLAS-common \
 RUN git clone https://github.com/ROCm/hipBLASLt
 RUN cd hipBLASLt \
     && git checkout ${HIPBLASLT_BRANCH} \
-    && ./install.sh -d --architecture ${PYTORCH_ROCM_ARCH} ${LEGACY_HIPBLASLT_OPTION} \
+    && apt-get install -y llvm-dev \
+    && ./install.sh -dc --architecture ${PYTORCH_ROCM_ARCH} ${LEGACY_HIPBLASLT_OPTION} \
     && cd build/release \
     && make package
 RUN mkdir -p /app/install && cp /app/hipBLASLt/build/release/*.deb /app/hipBLAS-common/build/*.deb /app/install
@@ -108,11 +111,24 @@ RUN git clone ${FA_REPO}
 RUN cd flash-attention \
     && git checkout ${FA_BRANCH} \
     && git submodule update --init \
-    && MAX_JOBS=64 GPU_ARCHS=${PYTORCH_ROCM_ARCH} python3 setup.py bdist_wheel --dist-dir=dist
+    && GPU_ARCHS=$(echo ${PYTORCH_ROCM_ARCH} | sed -e 's/;gfx1[0-9]\{3\}//g') python3 setup.py bdist_wheel --dist-dir=dist
 RUN mkdir -p /app/install && cp /app/pytorch/dist/*.whl /app/install \
     && cp /app/vision/dist/*.whl /app/install \
     && cp /app/flash-attention/dist/*.whl /app/install
 
+FROM base AS build_aiter
+ARG AITER_BRANCH
+ARG AITER_REPO
+RUN --mount=type=bind,from=build_pytorch,src=/app/install/,target=/install \
+    pip install /install/*.whl
+RUN git clone --recursive ${AITER_REPO}
+RUN cd aiter \
+    && git checkout ${AITER_BRANCH} \
+    && git submodule update --init --recursive \
+    && pip install -r requirements.txt
+RUN pip install pyyaml && cd aiter && PREBUILD_KERNELS=1 GPU_ARCHS=gfx942 python3 setup.py bdist_wheel --dist-dir=dist && ls /app/aiter/dist/*.whl
+RUN mkdir -p /app/install && cp /app/aiter/dist/*.whl /app/install
+
 FROM base AS final
 RUN --mount=type=bind,from=build_hipblaslt,src=/app/install/,target=/install \
     dpkg -i /install/*deb \
@@ -128,8 +144,11 @@ RUN --mount=type=bind,from=build_amdsmi,src=/app/install/,target=/install \
     pip install /install/*.whl
 RUN --mount=type=bind,from=build_pytorch,src=/app/install/,target=/install \
     pip install /install/*.whl
+RUN --mount=type=bind,from=build_aiter,src=/app/install/,target=/install \
+    pip install /install/*.whl
 
 ARG BASE_IMAGE
+ARG HIPBLAS_COMMON_BRANCH
 ARG HIPBLASLT_BRANCH
 ARG LEGACY_HIPBLASLT_OPTION
 ARG RCCL_BRANCH
@@ -142,6 +161,8 @@ ARG PYTORCH_REPO
 ARG PYTORCH_VISION_REPO
 ARG FA_BRANCH
 ARG FA_REPO
+ARG AITER_BRANCH
+ARG AITER_REPO
 RUN echo "BASE_IMAGE: ${BASE_IMAGE}" > /app/versions.txt \
     && echo "HIPBLAS_COMMON_BRANCH: ${HIPBLAS_COMMON_BRANCH}" >> /app/versions.txt \
     && echo "HIPBLASLT_BRANCH: ${HIPBLASLT_BRANCH}" >> /app/versions.txt \
@@ -155,4 +176,5 @@ RUN echo "BASE_IMAGE: ${BASE_IMAGE}" > /app/versions.txt \
     && echo "PYTORCH_REPO: ${PYTORCH_REPO}" >> /app/versions.txt \
     && echo "PYTORCH_VISION_REPO: ${PYTORCH_VISION_REPO}" >> /app/versions.txt \
     && echo "FA_BRANCH: ${FA_BRANCH}" >> /app/versions.txt \
-    && echo "FA_REPO: ${FA_REPO}" >> /app/versions.txt
+    && echo "AITER_BRANCH: ${AITER_BRANCH}" >> /app/versions.txt \
+    && echo "AITER_REPO: ${AITER_REPO}" >> /app/versions.txt